Helsinki University of Technology Laboratory of Acoustics and Audio Signal Processing

Espoo 1999

VIRTUAL ACOUSTICS AND 3-D SOUND IN

MULTIMEDIA SIGNAL PROCESSING

Jyri Huopaniemi

Report 53

Helsinki University of Technology Laboratory of Acoustics and Audio Signal Processing

Espoo 1999

Report 53

VIRTUAL ACOUSTICS AND 3-D SOUND IN

MULTIMEDIA SIGNAL PROCESSING

Jyri Huopaniemi
Dissertation for the degree of Doctor of Science in Technology to be presented with due permission for
public examination and debate in Auditorium S4, Department of Electrical and Communications
Engineering, Helsinki University of Technology (Espoo, Finland) on the 5th of November, 1999, at 12
o'clock noon.

Helsinki University of Technology

Deparment of Electrical and Communications Engineering
Laboratory of Acoustics and Audio Signal Processing
Teknillinen korkeakoulu
Shk- ja tietoliikennetekniikan osasto
Akustiikan ja nenksittelytekniikan laboratorio

Helsinki University of Technology

Laboratory of Acoustics and Audio Signal Processing
P.O.Box 3000
FIN-02015 HUT
Tel. +358 9 4511
Fax +358 9 460 224
E-mail lea.soderman@hut.fi

Jyri Huopaniemi
Cover picture of Marienkirche Erkki Rousku
ISBN 951-22-4706-2
ISSN 1456-6303

Libella Oy
Espoo, Finland 1999

Abstract
In this work, aspects in real-time modeling and synthesis of three-dimensional
sound in the context of digital audio, multimedia, and virtual environments are
studied. The concept of virtual acoustics is discussed, which includes models for
sound sources, room acoustics, and spatial hearing.
Real-time virtual acoustics modeling is carried out using a real-time parametric room impulse response rendering technique proposed by the author. The
algorithm uses a well-known time-domain hybrid model, in which the direct sound
and early energy are modeled using geometrical acoustics, and diuse late energy
is modeled using a recursive reverberator. Sound source directivity, and dynamic
room acoustical phenomena such as time-varying early re ections from boundary
materials and air absorption are incorporated in the image-source model using
low-order digital lter approximations. A novel technique for estimation of source
directivity based on the reciprocity method is introduced.
Optimization of binaural systems is discussed. Novel methods for head-related
transfer function approximation based on frequency warping and balanced model
truncation are proposed by the author. Objective and subjective methods for
estimating the quality of virtual source generation in headphone listening are
introduced. For objective analysis, a binaural auditory model is used, which
predicts well the expected performance of dierent lter designs. Subjective
listening tests have been carried out to compare design techniques. The results
show that binaural lter design based on auditory criteria results in more ecient
and more perceptually relevant processing. This is the main result of the thesis.
Crosstalk canceled loudspeaker reproduction of 3-D sound is reviewed. Stereo
widening and virtual loudspeaker algorithms are discussed. An ecient, highquality design method of virtual surround lters for playback of multichannel
audio on two loudspeakers based on a warped shuer structure is presented.
The performance of the method has been veried in subjective listening tests.
The methods and results presented in this thesis are applicable to digital audio
signal processing in general, and especially to implementation of virtual acoustic
environments in multimedia and virtual reality systems.
3-D sound, auralization, head-related transfer function, binaural
technology, crosstalk canceling, spatial hearing, digital signal processing, auditory
displays, virtual acoustics, virtual environments
Keywords:

Preface
This work is a result of research carried out at the Laboratory of Acoustics and
Audio Signal Processing of the Helsinki University of Technology (HUT) in Espoo,
Finland, during the years 1995-1998, at the Center for Computer Research in
Music and Acoustics (CCRMA) of Stanford University during the year 1998, and
nalized in 1998-1999 at Nokia Research Center in Helsinki, Finland. I am most
grateful to my supervisor, Professor Matti Karjalainen, for his continuous interest
and support in my work over the years, and for our productive collaboration in
all research areas of this thesis. I'm also very grateful to Prof. Julius O. Smith
III for inviting and supervising my research work at CCRMA and for many
intriguing discussions and collaboration on HRTF lter design. Furthermore, I'd
like to thank the pre-examiners of my thesis, Prof. Richard Duda (San Jose State
University) and Dr. Tapio Lahti (Insinooritoimisto Akukon) for positive feedback
and very valuable comments.
During my thesis work, I had close cooperation with many researchers. A
number of early publications on physical modeling I co-authored with Prof. Karjalainen and Dr. Vesa Valimaki formed the basis for my research into virtual
acoustics. I'd like to thank Matti and Vesa for very fruitful collaboration, and for
guidance, reading and commenting most of my publications, including this thesis.
The Digital Interactive Virtual Acoustics (DIVA) group, headed by Prof. Tapio
Takala (HUT Laboratory of Telecommunications Software and Multimedia), has
been a project which has in uenced my research greatly. I'd like to thank the
DIVA group, especially Mr. Lauri Savioja and Mr. Tapio Lokki for collaboration
on implementation issues and on many excellent research articles. I worked with
the 3-D sound group at HUT Acoustics Lab on many interesting topics in 19961998. I'd like to thank Ms. Riitta Vaananen for our productive collaboration
in MPEG-4 scene description, in the DIVA project, and in many publications.
Mr. Klaus Riederer was the key person in obtaining the HRTF data which is
the basis for many parts of this work. Mr. Martti Rahkila and I have collaborated on many internet and web related projects. I admire Mara's style and net
knowledge. Mr. Tero Tolonen's expertise in DSP, physical modeling and music
has been inspiring. I've enjoyed Dr. Unto Laine's wisdom in warping, audio DSP,
and philosophy. Mr. Ville Pulkki's ideas and our cooperation in binaural auditory
modeling have been very fruitful. Mr. Aki Harma and I have shared common in7

8
terests on research of frequency warping and its applications. Mr. Jussi Hynninen
created the ultimate listening test software, GP2, which was used in the virtual
surround test. Mr. Antti Jarvinen has introduced me to many interesting people,
and we've collaborated on many research issues. Ms. Lea Soderman has assisted
me in many practical issues over the years. I'm very grateful to these ne persons
and the entire HUT Acoustics Lab personnel for cooperation, friendships and an
inspiring working atmosphere during the ve years I spent there as a researcher.
My colleagues at Nokia Research Center (NRC), Speech and Audio Systems
Laboratory, are thanked for professional and academic support in this work. In
particular I'd like to thank Dr. Petri Haavisto, Mr. Mauri Vaananen, Mr. Matti
Hamalainen and Mr. Nick Zacharov. I've co-authored research papers with Nick
and Matti, and I've enjoyed Nick's expertise in subjective experiments and Matti's
knowledge in audio DSP. I'd also like to thank the entire NRC Speech and Audio
Systems Lab's 3-D Audio Group for hard-working research eorts and for making
a great working atmosphere. It has been a pleasure to collaborate with Mr. Juha
Backman (Nokia Mobile Phones) during many years. His knowledge of physics,
acoustics and audio is remarkable. I'd like to thank Dr. Durand Begault and Dr.
Elizabeth Wenzel (NASA Ames Research Center) for their continuing interest in
our virtual acoustics research, and for very inspiring discussions. Dr. Jonathan
Mackenzie is acknowledged for collaboration in HRTF modeling using the BMT
method. I've had the pleasure to work together with Mr. Eric Scheirer (MIT
Media Lab) on MPEG-4 audio scene description related work, and we've collaborated on two publications. Dr. Jean-Marc Jot (Creative Labs) and I have had
many very fruitful discussions on virtual acoustics, and we've worked together on
MPEG-4 audio rendering issues. I'd like to sincerely thank these great researchers
for their interest in my work.
There are also many friends and my family I'd like to acknowledge. Very
special thanks go to Mr. Saku Heiskanen, Ms. Kaisu Iisakkila, Mr. Martti Rahkila
and Ms. Outi Kosonen for friendship and great company. I've played music with
Saku in a number of bands for 15 years, and it continues to bring a lot of joy
into my life. The friends I lived with in Toukola, the HC Finland, the Munich
people, the Mielentila/Mimedia crew, the Lautta musicians are acknowledged for
making me do, see and hear other things than work every now and then. Most
importantly, I'd like to thank my girlfriend, Ms. Patty Huang, for everything,
including valuable comments to the thesis manuscript. Finally, I'd like to thank
my parents and their families for support all through my years of research. Special
thanks go to my mother, Ms. Sinikka Huopaniemi, Mr. Vesa Hatakka, and my
father, Mr. Hannu Huopaniemi, Ms. Annariitta Huopaniemi, and my brothers,
Mikko and Juha. The nancial support of Nokia Research Center, the Academy
of Finland, the Emil Aaltonen Foundation, the Helsinki University of Technology
Foundation, and the Nokia Foundation is greatly acknowledged.
Lauttasaari, Helsinki, October 13, 1999
Jyri Huopaniemi

Table of Contents
1

Introduction

1.1
1.2
1.3
1.4

Background . . . . . . . . .
Overview of the Thesis . . .
Contributions of the Author
Related Publications . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

Virtual Acoustics

2.1 Virtual Acoustic Modeling Concepts . . . . . . . . . . . .

2.2 Methods for 3-D Sound Reproduction . . . . . . . . . . . .
2.2.1 Headphone Reproduction . . . . . . . . . . . . . . .
2.2.2 Loudspeaker Reproduction . . . . . . . . . . . . . .
2.2.3 Multichannel Techniques . . . . . . . . . . . . . . .
2.3 Implementation of Virtual Acoustic Systems . . . . . . . .
2.3.1 Real-Time Geometrical Room Acoustics Approach .
2.4 DIVA System . . . . . . . . . . . . . . . . . . . . . . . . .
2.5 Virtual Acoustics in Object-Based
Multimedia . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . .
Sound Source Modeling

3.1 Properties of Sound Sources . . . . . . . . . . . . . . .

3.1.1 On the Directivity of Musical Instruments . . .
3.1.2 Source Directivity Models . . . . . . . . . . . .
3.1.3 Directional Filtering . . . . . . . . . . . . . . .
3.1.4 Set of Elementary Sources . . . . . . . . . . . .
3.2 Sound Radiation from the Mouth . . . . . . . . . . . .
3.2.1 Analytical Formulation of Head Directivity . . .
3.2.2 Measurement of Head Directivity . . . . . . . .
3.2.3 Modeling and Measurement Result Comparison
3.3 Discussion and Conclusions . . . . . . . . . . . . . . .
9

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

19

19
22
23
24

27

27
30
30
30
31
32
34
37

. . . . 40
. . . . 41
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

43

45
45
45
46
47
48
49
50
52
53

10

TABLE OF CONTENTS

Enhanced Geometrical Room Acoustics Modeling

.
.
.
.

.
.
.
.

.
.
.
.

57

Binaural Modeling and Reproduction

5.1 Properties and Modeling of HRTFs . . . . . . . . . . . . . . .

5.1.1 Interaural Time and Phase Dierence . . . . . . . . . .
5.1.2 Interaural Level Dierence and Spectral Cues . . . . .
5.1.3 Other Cues for Sound Localization . . . . . . . . . . .
5.1.4 Distance-Dependency of the HRTF . . . . . . . . . . .
5.1.5 Functional Modeling of HRTFs . . . . . . . . . . . . .
5.2 Auditory Analysis of HRTFs and Application to Filter Design
5.2.1 Properties of the Human Peripheral Hearing . . . . . .
5.2.2 Auditory Smoothing . . . . . . . . . . . . . . . . . . .
5.2.3 Auditory Weighting . . . . . . . . . . . . . . . . . . . .
5.2.4 Frequency Warping . . . . . . . . . . . . . . . . . . . .
5.3 Digital Filter Design of HRTFs . . . . . . . . . . . . . . . . .
5.3.1 HRTF Preprocessing . . . . . . . . . . . . . . . . . . .
5.3.2 Error Norms . . . . . . . . . . . . . . . . . . . . . . . .
5.3.3 Finite Impulse-Response Methods . . . . . . . . . . . .
5.3.4 Innite Impulse-Response Methods . . . . . . . . . . .
5.3.5 Warped Filters . . . . . . . . . . . . . . . . . . . . . .
5.4 Binaural System Implementation . . . . . . . . . . . . . . . .
5.4.1 Interpolation and Commutation of HRTF Filters . . .
5.5 Objective and Subjective Evaluation of
HRTF Filter Design Methods . . . . . . . . . . . . . . . . . .
5.5.1 Objective Methods . . . . . . . . . . . . . . . . . . . .
5.5.2 Subjective Methods . . . . . . . . . . . . . . . . . . . .
5.6 Experiments in Binaural Filter Design and Evaluation . . . . .
5.6.1 Binaural Filter Design Experiment . . . . . . . . . . .
5.6.2 Evaluation of Non-Individualized HRTF
Performance . . . . . . . . . . . . . . . . . . . . . . . .
5.6.3 Evaluation of Individualized HRTF Performance . . . .
5.6.4 Eects of HRTF Preprocessing and Equalization . . . .
5.7 Discussion and Conclusions . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

67

67
71
77
78
79
85
86
86
87
88
88
90
92
95
96
97
100
102
104
106
106
109
109
109
112
118
131
134

Crosstalk Canceled Binaural Reproduction

.
.
.
.

.
.
.
.

139

4.1
4.2
4.3
4.4

Acoustical Material Filters . . . . . . . . . . . . .

Air Absorption Filters . . . . . . . . . . . . . . .
Implementation of Extended Image-Source Model
Discussion and Conclusions . . . . . . . . . . . .

6.1 Theory of Crosstalk Canceling . . . . . . . . .

6.1.1 Symmetrical Crosstalk Canceling . . .
6.1.2 Asymmetrical Crosstalk Canceling . .
6.1.3 Other Crosstalk Canceling Structures .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

58
62
63
64

140
141
144
145

TABLE OF CONTENTS

6.2 Virtual Source Synthesis . . . . . . . . . . . . . . . . . . . . . . .

6.2.1 Symmetrical Listening Position . . . . . . . . . . . . . . .
6.2.2 Decorrelation of Virtual Sources . . . . . . . . . . . . . . .
6.2.3 Asymmetrical Listening Position . . . . . . . . . . . . . .
6.2.4 Binaural and Crosstalk Canceled Binaural Conversion Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.5 Virtual Center Channel . . . . . . . . . . . . . . . . . . . .
6.3 Stereophonic Image Widening . . . . . . . . . . . . . . . . . . . .
6.3.1 Traditional Stereophonic Image Enhancement Techniques .
6.3.2 Stereo Widening Based on 3-D Sound Processing . . . . .
6.4 Virtual Loudspeaker Filter Design . . . . . . . . . . . . . . . . . .
6.4.1 Finite Impulse-Response Methods . . . . . . . . . . . . . .
6.4.2 Innite Impulse-Response Methods . . . . . . . . . . . . .
6.4.3 Conclusion and Discussion . . . . . . . . . . . . . . . . . .
6.5 System Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5.1 Widening the Listening Area . . . . . . . . . . . . . . . . .
6.5.2 Analysis of Crosstalk Canceled Binaural Designs . . . . . .
6.6 Subjective Evaluation of Virtual Surround Systems . . . . . . . .
6.6.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . .
6.6.2 Experimental Procedure and Setup . . . . . . . . . . . . .
6.6.3 Test Items and Grading . . . . . . . . . . . . . . . . . . .
6.6.4 Results and Discussion . . . . . . . . . . . . . . . . . . . .
6.7 Discussion and Conclusions . . . . . . . . . . . . . . . . . . . . .
7

Conclusions

Bibliography

11
145
145
148
149
149
149
151
151
153
154
156
156
156
157
157
159
161
161
162
164
165
165
167
169

12

TABLE OF CONTENTS

List of Symbols
a
 (! )
c
c(n)
C (z )
fCE
fCB
f
fs
fro
frn
Fk (z )
g
h
h(n)
hA ; : : : ; hE
hm (kr)
h0m (ka)
H (ej! )

n
pr
Pm

r
RT60
R(j! )
0
S
t
T
Tk (z )

radius of human head; attenuation in dB per meter

absorption coecient
speed of sound in air
cepstrum
crosstalk canceling lter
ERB scale frequency bandwidth
Bark scale frequency bandwidth
temporal frequency
sampling frequency
oxygen relaxation frequency
nitrogen relaxation frequency
listener model lter block
amplication factor
molar concentration of water vapor
impulse response
HRTF lter coecients
spherical Hankel function of order m in kr
derivative of spherical Hankel function of order m with respect to ka
frequency response
warping factor
integer variable
reference ambient atmospheric pressure
Legendre polynomial of order m
elevation angle
distance from source to listener
reverberation time
re ectance
density of air
sound source
time variable
sampling interval
auralization lter block
13

14

TABLE OF CONTENTS

T
T0
high
group
low

e
!
x(n)
y (n)
z
Z (j! )

temperature in Kelvin
reference air temperature
high-frequency interaural time delay
group-delay interaural time delay
low-frequency interaural time delay
azimuth angle
azimuth angular error
angular frequency
input signal
output signal
z -transform variable
normalized impedance

15

16

TABLE OF CONTENTS

List of Abbreviations
ANCOVA
ANOVA
AR
ARMA
BEM
BIFS
BRIR
BMT
CCRMA
CD
CF
CT
DF
DFT
DIVA
DRIR
DSP
DTF
DVD
ERB
FDTD
FEM
FFT
FIR
FM
GTFB
GUI
HOA
HRIR
HRTF
HSV
HUT
IACC
IFFT
IIR

analysis of covariance
analysis of variance
auto-regressive
auto-regressive moving average
boundary element method
binary format for scenes
binaural room impulse response
balanced model truncation
Center for Computer Research in Music and Acoustics
compact disc; committee draft
Caratheodory-Fejer
cross-talk
direct form
discrete Fourier transform
digital interactive virtual acoustics
direct room impulse response
digital signal processing
directional transfer function
digital versatile disc
equivalent rectangular bandwidth
nite dierence time domain
nite element method
fast Fourier transform
nite impulse response
frequency modulation
gammatone lterbank
graphical user interface
Hankel norm optimal approximation
head-related impulse response
head-related transfer function
Hankel singular values
Helsinki University of Technology
interaural cross-correlation
inverse fast Fourier transform
innite impulse response

TABLE OF CONTENTS

ILD
IPD
IRCAM
ISO
ITD
JND
KLE
LFE
LMS
LS
LTI
MA
MPEG
PCA
PRIR
RIR
SEA
SER
SPL
SPSS
STFT
SVD
TAFC
TTS
VAD
VHT
VL
VRML
VS
WFIR
WIIR
WLS
WSF

interaural level dierence

interaural phase dierence
Institut de Recherche et Coordination Acoustique/Musique
international standardization organization
interaural time dierence
just noticeable dierence
Karhunen-Loeve expansion
low frequency energy
least-mean square
least squares
linear time-invariant
moving average
moving picture experts group
principal components analysis
parametric room impulse response
room impulse response
statistical energy analysis
signal-to-error ratio
sound pressure level
Statistical Package for Social Sciences
short-time Fourier transform
singular value decomposition
two alternatives forced choice
text-to-speech synthesis
virtual auditory display
virtual home theatre
virtual loudspeaker
virtual reality modeling language
virtual speaker
warped nite impulse response
warped innite impulse response
weighted least squares
warped shuer lter

17

18

TABLE OF CONTENTS

Chapter 1
Introduction
The topic of this thesis is physically-based modeling of room acoustical systems
and human spatial hearing by means of digital signal processing (DSP). A concept of virtual acoustics is explored, which deals with three major subsystems in
acoustical communication as shown in Fig. 1.1: source modeling, transmission
medium modeling, and listener modeling (Begault, 1994, p. 6). Underlying DSP
concepts for ecient and auditorily relevant multimedia sound processing are
studied, and a general framework for virtual acoustics modeling is proposed. The
goal in this thesis is to develop technology for delivering an acoustical message
in a virtual reality system from the source to the receiver as it would happen in
a real-world situation.
1.1

Background

The characteristics of human hearing form the source of information for the research of 3-D sound and virtual acoustics (Blauert, 1997). Within this framework,
S o u rc e

E n v ir o n m e n t

L is te n e r

Figure 1.1: The concept of virtual acoustics comprises source modeling, environment modeling, and listener modeling.
19

20

1. Introduction

Term

Explanation

Monotic or Monaural
stimulus presented to only one ear
Dichotic
stimuli that are dierent at the two ears
Diotic
stimuli that are identical at the ears
Binaural
any stimulus that is diotic or dichotic
Table 1.1: Terminology of binaural technology (Yost and Gourevitch, 1987, p.
50).
distinctive terms describing methods for spatial localization and processing of
sounds have been dened. Fundamental denitions of binaural terminology are
presented in Table 1.1 (Yost and Gourevitch, 1987). In this work, the term
binaural is constrained to mean only such processing that a three-dimensional
auditory illusion is created for headphone listening (Mller, 1992). The term
crosstalk canceled binaural processing is used to refer to manipulation of binaural information for listening with two loudspeakers 1 . The term auralization is
understood as rendering audible a modeled acoustical space in such a way that
the three-dimensional sound illusion is preserved (Kleiner et al., 1993). By virtual
acoustics the concept of auralization is expanded to include virtual reality aspects
such as dynamic movement, dynamic source and acoustic space characterization,
and immersion using advanced control systems (Takala et al., 1996; Savioja et
al., 1999). Auditory display is a method or device used to display information
to human observers via the auditory system (Shinn-Cunningham et al., 1997, p.
611).
Sound has many roles in a multimedia or virtual reality system. In most cases,
sound is used primarily for communicating (informational, navigational speech
and non-speech audio) or entertaining (music, voice, ambience) purposes. This
thesis deals with the processing and manipulation of any input sound, modifying
it so that the outcome resembles a physical or articial three-dimensional auditory environment. There is a long history of research on spatial hearing and
three-dimensional sound (Blauert, 1997). Lord Rayleigh (J.W. Strutt, 3rd Baron
of Rayleigh) presented the rst in uential theory of spatial hearing, called the
\duplex theory," in the turn of the 20th century (Rayleigh, 1907). The rst virtual acoustic concepts were explored and examined in the eld of teleoperation.
Spandock (1934) designed scale models and auralized concert hall designs in the
1930's 2. Even prior to that, during World War I, enemy aircraft detection was
enhanced by imitating spatial hearing cues for accurate three-dimensional source
localization using a pseudophone (magnied ear canal and pinnae) 3 . Both of
1

The term transaural that refers to same type of processing is trademarked by Cooper and
Bauck (1989). See (Gardner, 1998a) for review of crosstalk canceling technology.
2 See (Kleiner et al., 1993; Kuttru, 1993; Begault, 1994; Blauert, 1997) for review of
auralization and spatial hearing.
3 See (Wenzel, 1992; Shinn-Cunningham et al., 1997) for review of auditory displays.

1.1

Background

21

these early applications indeed had the same major goals as virtual acoustics
has today: virtual design of audiovisual spaces, i.e., exploration of acoustical behavior virtually, and accurate three-dimensional simulation and reproduction of
auditory events. In other words, the aim is to augment the hearing sense when
using unideal reproduction devices such as headphones or a pair of loudspeakers.
With the rapid advance of multimedia technology, the virtual acoustics concept
has in recent years expanded to cover even more areas, ranging from physical
modeling of sound sources, via room acoustics rendering, to modeling of spatial
hearing cues.
The application area of virtual acoustic technology is wide, ranging from entertainment and telecommunications to virtual reality and acoustics simulation.
Three-dimensional audio systems are often associated with virtual acoustic displays (VAD). The VADs play an important role in simulations and the virtual
reality technology, where source, environment, and listener models are combined
to produce an immersive experience (Begault, 1994). Game and multimedia
technology are today the leading edge in consumer applications of 3-D sound.
Teleconferencing and videoconferencing are another popular application area utilizing three-dimensional sound processing. Virtual environment technology, with
standards like Virtual Reality Modeling Language (VRML) (ISO/IEC, 1997) and
MPEG-4 (ISO/IEC, 1999), is further expanding the use of virtual acoustics in
the context of audiovisual simulation. The growth of positional 3-D sound technology and applications that use binaural headphone or loudspeaker processing
has been signicant in the last few years, especially with the emergence of multichannel audio formats. Multichannel audio, for example in DVD movies, is often
played back in domestic environments using only two channels. Therefore, the
concepts of virtual surround (or virtual home theater, VHT) for headphone and
loudspeaker reproduction have been developed to enhance the three-dimensional
listening experience.
From the viewpoint of sound reproduction, the research conducted in this thesis also involves improving the capabilities of traditional two-channel stereophony.
Stereophonic system implementation can be divided into three stages: sound
pickup (microphone), sound mixing (processing), and sound reproduction (Theile,
1991). This research is focused on the last two stages of stereophonic reproduction. Virtual acoustic technology and binaural processing for headphone
and loudspeaker listening has produced a new era of virtual audio reality to
two-channel stereophony. With the aid of modern signal processing methods
rewiewed and presented in this thesis, it is possible to design and implement
three-dimensional virtual auditory environments and audio eects for consumer
audiovisual products used in computer, television and hi- systems.

22
1.2

1. Introduction

Overview of the Thesis

The topics covered in this thesis can be considered interdisciplinary. They range
from psychoacoustics and spatial hearing to digital signal processing aspects and
implementation, and from room acoustics and the behavior of musical instruments to real-time system design and specication. The aim of this thesis is not
to seamlessly cover all the interesting and unexplored research problems in the
cross-section of all the areas in virtual acoustics. Instead, this thesis is written
more in the spirit of presenting \selected topics in virtual acoustics," concentrating on the new innovations and contributions by the author related to the
eld. Some of the underlying theory is discussed in great detail, especially that of
spatial hearing and manipulation of head-related transfer functions, to produce a
complete view of the problem. Some background topics (for example, geometrical
room acoustics modeling and implementation) are, however, only brie y introduced to the reader. Therefore a moderate to good background knowledge of the
topic area is expected.
The scope of this work has been twofold. In the rst part, topics in source
and room acoustics modeling related to real-time processing of virtual acoustic
environments are covered (Savioja et al., 1999). New methods are presented for
source and environment modeling and lter design. In the second part, headrelated transfer function (HRTF) modeling, approximation, and evaluation techniques for headphone and loudspeaker reproduction are presented. The work
reviews existing theories and methods for realization of a real-time auralization
system, and presents new methodology, lter design and implementation issues
of HRTFs and virtual acoustics. Furthermore, novel methods and results of subjective and objective evaluation of 3-D audio systems for both headphone and
loudspeaker listening are presented. The author motivates the use of auditory
criteria in HRTF-based lter design for binaural and crosstalk canceled binaural
processing. The results are veried by carefully conducted subjective listening
experiments, and a binaural auditory model is used for predicting further lter design and implementation performance. One result of this thesis has been
the author's contribution to the auralization specication and implementation of
the DIVA (Digital Interactive Virtual Acoustics) Virtual Audio Reality System
(Takala et al., 1996; Savioja et al., 1999) developed at the Helsinki University of
Technology. The DIVA system is a real-time virtual concert hall performance environment including MIDI-to-movement mapping, animation, physical modeling,
room acoustics modeling, and 3-D sound simulation. Furthermore, the author
has actively contributed to the MPEG-4 version 2 standardization in the eld
of virtual acoustic parametrization (Swaminathan et al., 1999; Vaananen and
Huopaniemi, 1999; Scheirer et al., 1999). The guidelines for HRTF lter design
provided by the author in this thesis give a scientically solid framework for optimized reproduction of 3-D sound in a wide range of applications such as virtual
reality, teleconferencing, and professional and domestic audio systems.

1.3

Contributions of the Author

23

This thesis is divided as follows. An overview of virtual acoustic technology

is given in Chapter 2. In Chapter 3, topics in sound source parametrization
and modeling are studied, and new methods for source modeling are presented.
Chapter 4 deals with real-time room acoustical issues, concentrating on novel
lter design aspects for phenomena such as air absorption and re ections from
room boundaries. Chapter 5 is devoted to HRTF approximation and lter design for measurements at both close distance and far distance, and on dummy
heads as well as human subjects. Optimization of binaural processing for headphone listening is discussed, and new methods and results for subjective and
objective evaluation of binaural lter design are presented. In Chapter 6, binaural techniques are extended for loudspeaker reproduction by crosstalk canceling
mechanisms and virtual loudspeaker technology. A new algorithm and lter design for stereo widening and virtual surround processing is presented. Subjective
evaluation has been carried out to evaluate the quality of the designed virtual
surround lters. Finally, conclusions are drawn and directions for further work
are given in Chapter 7.
1.3

Contributions of the Author

The author's contributions to the eld of 3-D audio and virtual acoustic technology, presented in this thesis, can be summarized as follows. In virtual acoustic
environment related issues (Chapter 2), the author has participated in the DIVA
group research and contributed to system architecture denition and both nonreal-time and real-time auralization lter design and implementation (Huopaniemi
et al., 1994; Takala et al., 1996; Huopaniemi et al., 1996; Savioja et al., 1997,
1999). The work in the DIVA group has been a result of long cooperation with
Mr. Lauri Savioja, Ms. Riitta Vaananen, and Mr. Tapio Lokki. The author
has also actively contributed to MPEG-4 standardization work and introduced
parametrization strategies for physically-based virtual acoustic rendering (Swaminathan et al., 1999; Scheirer et al., 1999; Vaananen and Huopaniemi, 1999).
In source modeling (Chapter 3), the author investigated and was the rst
to introduce the use of physical musical instrument models as sound sources in
a virtual acoustic simulation environment (Huopaniemi et al., 1994). Concepts
and modeling strategies for sound source directivity have been presented by the
author in (Huopaniemi et al., 1994, 1995; Karjalainen et al., 1995; Valimaki et
al., 1996; Huopaniemi et al., 1999a). This work has been a result of fruitful
cooperation with Prof. Matti Karjalainen and Dr. Vesa Valimaki. A novel method
for measuring and modeling sound radiation from the mouth has been suggested
by the author in (Huopaniemi et al., 1999a).
In room acoustics modeling (Chapter 4), the author has contributed to lter design and real-time simulation of boundary re ections and air absorption
(Huopaniemi et al., 1997) as well as reverberation algorithm design (Huopaniemi

24

1. Introduction

et al., 1994; Vaananen et al., 1997).

The author's contribution to research in HRTF lter design and binaural
systems (Chapters 5 and 6) has resulted in new binaural lter design methods based on frequency warping (Huopaniemi and Karjalainen, 1996a,b, 1997)
and balanced model truncation (Mackenzie et al., 1997) (in cooperation with
Prof. Karjalainen, Dr. Valimaki, and Dr. Jonathan Mackenzie). The author
has also investigated distance-dependency in HRTFs (Huopaniemi and Riederer,
1998). Furthermore, objective and subjective evaluation of auditory binaural audio signal processing has been performed (Huopaniemi and Smith, 1999; Pulkki et
al., 1999; Huopaniemi et al., 1999b) and practical implementations (Karjalainen
et al., 1999; Zacharov et al., 1999; Zacharov and Huopaniemi, 1999) have been
created by the author.
Practical and useful results of the thesis work include methods for real-time
virtual acoustics rendering, virtual loudspeaker design, enhanced stereo imaging,
and widening of the listening area in two-channel loudspeaker reproduction.
1.4

Related Publications

During the research work, the author has contributed to the following publications, parts of which have been used in this thesis.
Reviewed Journal Articles

Savioja, L., Huopaniemi, J., Lokki, T. and Vaananen, R. 1999. Creating interactive
virtual acoustic environments, Journal of the Audio Engineering Society 47(9): 675{
705, September 1999.
Scheirer, E., Vaananen, R., Huopaniemi, J. 1999. AudioBIFS: The MPEG-4
standard for audio eects processing, IEEE Transactions on Multimedia 1(3): 237{
250, September 1999.
Huopaniemi, J., Zacharov, N., and Karjalainen, M. 1999b. Objective and subjective evaluation of head-related transfer function lter design, Journal of the Audio
Engineering Society 47(4): 218{239, April 1999.
Pulkki, V., Karjalainen, M. and Huopaniemi, J. 1999. Analyzing virtual sound
sources using a binaural auditory model, Journal of the Audio Engineering Society
47(4): 203{217, April 1999.
Mackenzie, J., Huopaniemi, J., Valimaki, V. and Kale, I. 1997. Low-order modelling of head-related transfer functions using balanced model truncation, IEEE Signal
Processing Letters 4(2): 39{41.
Valimaki, V., Huopaniemi, J., Karjalainen, M. and Janosy, Z. 1996. Physical
modeling of plucked string instruments with application to real-time sound synthesis,
Journal of the Audio Engineering Society 44(5): 331{353.

1.4

Related Publications

25

International Conference Articles

Zacharov, N., and Huopaniemi, J. 1999. Results of a round robin subjective evaluation
of virtual home theatre sound systems, Presented at the 107th Convention of the Audio
Engineering Society, preprint 5760, New York.
Huopaniemi, J., Kettunen, K. and Rahkonen, J. 1999. Measurement and modeling
techniques for directional sound radiation from the mouth, Proceedings of the 1999
IEEE Workshop of Applications of Signal Processing to Audio and Acoustics, Mohonk
Mountain House, New Paltz, New York, pp. 183{186.
Huopaniemi, J. and Smith, J. O. 1999. Spectral and time-domain preprocessing
and the choice of modeling error criteria in binaural digital lter design, Proceedings
of the 16th Audio Engineering Society International Conference, Rovaniemi, Finland,
April 10-12, 1999, pp. 301{312.
Zacharov, N., Hamalainen, M. and Huopaniemi, J. 1999. Round robin subjective evaluation of virtual home theatre systems at the AES 16th international conference, Proceedings of the 16th Audio Engineering Society International Conference,
Rovaniemi, Finland, April 10-12, 1999, pp. 544{556.
Vaananen, R. and Huopaniemi, J. 1999. Spatial presentation of sound in scene
description languages, Presented at the 106th Convention of the Audio Engineering
Society, preprint 4921, Munich, Germany.
Huopaniemi, J., Zacharov, N. and Karjalainen, M. 1998. Objective and subjective evaluation of head-related transfer function lter design, Presented at the 105th
Convention of the Audio Engineering Society, preprint 4805, San Francisco, CA, USA
(invited paper).
Huopaniemi, J. and Riederer, K. 1998. Measuring and modeling the eect of distance in head-related transfer functions, Proceedings of the joint meeting of the Acoustical Society of America and the International Congress on Acoustics, Washington, USA,
pp. 2083{2084.
Huopaniemi, J., Savioja, L. and Karjalainen, M. 1997. Modeling of re ections
and air absorption in acoustical spaces: a digital lter design approach, Proceedings of
the 1997 IEEE Workshop of Applications of Signal Processing to Audio and Acoustics,
Mohonk Mountain House, New Paltz, New York.
Karjalainen, M., Harma, A., Laine, U. and Huopaniemi, J. 1997. Warped lters
and their audio applications, Proceedings of the 1997 IEEE Workshop of Applications
of Signal Processing to Audio and Acoustics, Mohonk Mountain House, New Paltz,
New York.
Vaananen, R., Valimaki, V., Huopaniemi, J. and Karjalainen, M. 1997. Ecient
and parametric reverberator for room acoustics modeling, Proceedings of the International Computer Music Conference, Thessaloniki, Greece, pp. 200{203.
Hiipakka, D. G. J., Hanninen, R., Ilmonen, T., Napari, H., Lokki, T., Savioja, L.,
Huopaniemi, J., Karjalainen, M., Tolonen, T., Valimaki, V., Valimaki, S. and Takala,
T. 1997. Virtual Orchestra Performance, Visual Proceedings of SIGGRAPH'97, ACM
SIGGRAPH, Los Angeles, p. 81.
Huopaniemi, J. and Karjalainen, M. 1997. Review of digital lter design and
implementation methods for 3-D sound, Presented at the 102nd Convention of the
Audio Engineering Society, preprint 4461, Munich, Germany.

26

1. Introduction

Huopaniemi, J., Savioja, L. and Takala, T. 1996. DIVA virtual audio reality
system, Proceedings of the International Conference on Auditory Display, Palo Alto,
California, USA.
Takala, T., Hanninen, R., Valimaki, V., Savioja, L., Huopaniemi, J., Huotilainen,
T. and Karjalainen, M. 1996. An integrated system for virtual audio reality, Presented
at the 100th Convention of the Audio Engineering Society, preprint 4229, Copenhagen,
Denmark.
Huopaniemi, J. and Karjalainen, M. 1996a. Comparison of digital lter design
methods for 3-D sound, Proceedings of the IEEE Nordic Signal Processing Symposium,
Espoo, Finland, pp. 131{134.
Huopaniemi, J. and Karjalainen, M. 1996b. HRTF lter design based on auditory
criteria, Proceedings of the Nordic Acoustical Meeting (NAM'96), Helsinki, Finland,
pp. 323{330.
Karjalainen, M., Huopaniemi, J. and Valimaki, V. 1995. Direction-dependent
physical modeling of musical instruments, Proceedings of the International Congress
on Acoustics, Vol. 3, Trondheim, Norway, pp. 451{454.
Huopaniemi, J., Karjalainen, M. and Valimaki, V. 1995. Physical models of musical
instruments in real-time binaural room simulation, Proceedings of the International
Congress on Acoustics, Trondheim, Norway, pp. 447{450.
Huopaniemi, J., Karjalainen, M., Valimaki, V. and Huotilainen, T. 1994. Virtual
instruments in virtual rooms { a real-time binaural room simulation environment for
physical modeling of musical instruments, Proceedings of the International Computer
Music Conference, Aarhus, Denmark, pp. 455{462.

Parts of this present work have been published in the same or modied form in:
Huopaniemi, J. 1997.

Modeling of Human Spatial Hearing in the Context of Digital

, Helsinki University of Technology, Department of

Electrical and Communications Engineering, Laboratory of Acoustics and Audio Signal
Processing, p. 101. Lic.Tech. Thesis.
Audio and Virtual Environments

Chapter 2
Virtual Acoustics
Virtual acoustics is a general term for the modeling of acoustical phenomena and

systems with the aid of a computer. It may comprise many dierent elds of
acoustics, but in the context of this work the term is restricted to describe a
system ranging from sound source and acoustics modeling in rooms to spatial
auditory perception simulation in humans.
The virtual acoustic concept is closely related to other media for many reasons. The audiovisual technology (audio, still picture, video, animation etc.) is
rapidly integrating into a single interactive media. Research on audiovisual media modeling has increased dramatically in the last decade. Standardization of
advanced multimedia and virtual reality rendering and denition has been carried out for several years. Such examples are the Moving Pictures Expert Group
MPEG-4 (ISO/IEC, 1999) and the Virtual Reality Modeling Language (VRML)
(ISO/IEC, 1997) standards. The progress of multimedia has also introduced new
elds of research and application into audio and acoustics, one of which is virtual
acoustic environments1 and their relation to graphical presentations.
In this chapter, an overview of virtual acoustics modeling concepts is given.
The system that has been designed in the course of this study is outlined.
2.1

Virtual Acoustic Modeling Concepts

The design and implementation of virtual acoustic displays (VADs) can be divided
into three dierent stages as depicted in Fig. 2.1 (Huopaniemi, 1997; Savioja et
al., 1999): a) denition, b) modeling, and c) reproduction. The denition of a
virtual acoustic environment includes the prior knowledge of the system to be
implemented, that is, information about the sound sources, the room geometry,
and the listeners. The denition part is normally carried out o-line prior to the
1

Virtual acoustic environments are also called virtual auditory displays, auditory virtual
environments, virtual auditory space (Begault, 1994; Shinn-Cunningham et al., 1997).

27

28

2. Virtual Acoustics

DEFINITION

Source
Definition

Room Geometry
and Properties

HRTF
Database

SOURCE

MODELING

Source
Modeling
- natural audio
- synthetic audio
speech and sound
synthesis
- source directivity

MEDIUM
Room
Modeling
- modeling of
acoustic spaces
propagation
absorption
- artificial reverb

.
.

RECEIVER
Listener
Modeling
- modeling of
spatial hearing
HRTFs
simple models

.
.

REPRODUCTION
Multichannel

Binaural
headphone /
loudspeaker

Figure 2.1: Modeling concept of virtual acoustics (Huopaniemi, 1997; Savioja et

al., 1999).
real-time simulation process. In the modeling part the rendering is divided into
three tasks (Wenzel, 1992; Begault, 1994):
 Source Modeling
 Transmission Medium (Room Acoustics) Modeling
 Receiver (Listener) Modeling
This division is also often called the source-medium-receiver concept (Wenzel,
1992; Begault, 1994), which is widely known in many communication systems2 .
Interactive virtual acoustic environments and binaural room simulation systems
have been studied by, e.g., (Foster et al., 1991; Wenzel, 1992; Begault, 1994;
Shinn-Cunningham et al., 1997; Jot, 1999; Lehnert and Blauert, 1992; Vian and
Martin, 1992; Martin et al., 1993), but until recently the physical relevance and
the relation of acoustical and visual content has not been of major interest. The
term auralization (Kleiner et al., 1993) is understood as a subset of virtual acoustics referring to modeling and reproduction of sound elds (shown in Fig. 2.1).
2

In communication technology, the concept is often called source-channel-receiver.

2.1

Virtual Acoustic Modeling Concepts

29

Source modeling (Fig. 2.1) consists of methods which produce (and possibly
add physical character to) sound in an audiovisual scene. The most straightforward method is to use pre-recorded digital audio. In most auralization systems
sound sources are treated as omnidirectional point sources. This approximation is valid for many cases, but more accurate methods are, however, often
needed. For example, most musical instruments have radiation patterns which
are frequency-dependent. Typically sound sources radiate more energy to the
frontal hemisphere whereas sound radiation is attenuated and lowpass ltered
when the angular distance from the on-axis direction increases. In Chapter 3,
methods for ecient modeling of sound source directivity are presented.
The task of modeling sound propagation behavior in acoustical spaces (Fig. 2.1)
is an integral part of virtual acoustic simulation. The goal in most room acoustic
simulations has been to compute an energy time curve (ETC) of a room (squared
room impulse response) and, based on that, to derive room acoustical attributes
such as reverberation time (RT60 ). The ray-based methods, the ray tracing
(Krokstad et al., 1968; Kulowski, 1985) and the image-source methods (Allen
and Berkley, 1979; Borish, 1984), are the most often used modeling techniques.
Recently computationally more demanding wave-based techniques like nite element method (FEM), boundary element method (BEM), and nite-dierence
time-domain (FDTD) methods have also gained interest (Kuttru, 1995; Botteldooren, 1995; Savioja et al., 1995; Savioja, 1999). These techniques are, however,
suitable only for low-frequency simulation (Kleiner et al., 1993). In real-time
auralization, the limited computational capacity calls for simplications, modeling only the direct sound and early re ections using geometrical acoustics and
rendering the late reverberation by recursive digital lter structures (Schroeder,
1962; Moorer, 1979; Jot, 1999; Gardner, 1998b; Savioja et al., 1999).
In listener modeling (Fig. 2.1), the properties of human spatial hearing are
considered. Simple means for giving a directional sensation of sound are the
interaural level and time dierences (ILD and ITD), but they cannot resolve the
front-back confusion3. The head-related transfer function (HRTF) that models
the re ections and ltering by the head, shoulders and pinnae of the listener,
has been studied extensively during the past decade. With the development
of HRTF measurement techniques (Wightman and Kistler, 1989; Mller et al.,
1995) and ecient lter design methods (Kendall and Rodgers, 1982; Huopaniemi
et al., 1999b) real-time HRTF-based 3-D sound implementations have become
applicable in virtual environments.
3

See (Begault, 1994; Blauert, 1997; Mller, 1992; Kleiner et al., 1993; Kendall, 1995) for
fundamentals on spatial hearing and auralization.

30
2.2

2. Virtual Acoustics

Methods for 3-D Sound Reproduction

The illusion of three-dimensional sound elds can be created using various methods. The goal in 3-D sound simulation is to recreate any static or dynamic natural
or imaginary sound eld using a desired amount of transducers and proper signal
processing techniques. In this work, three methods have been considered:
 binaural reproduction ,
 crosstalk canceled binaural reproduction ,
 multichannel reproduction .
In order to more thoroughly understand the dierent aspects involved in 3-D
sound processing, the two binaural choices for implementation are rst considered,
namely binaural processing for headphone and loudspeaker listening. Figure 2.2
illustrates the dierences between these two processing methods.
2.2.1 Headphone Reproduction
In the case of binaural lter design, the HRTF measurement may directly be
approximated by various lter design methods (see Chapter 5) provided that
proper equalization is carried out. In the example of Fig. 2.2a, a monophonic
time-domain signal xm (n) is ltered with two HRTF lter approximations Hl(z)
and Hr (z) to create an image of a single virtual source. Known advantages of
binaural reproduction are the (trivial) facts that the acoustics of the listening
room and the positioning of the listener do not aect the perception. On the
other hand, head-tracking is required to compensate for head movements. Furthermore, it has been found that individual HRTFs should be used to create
the best performance in localization and externalization (Wenzel et al., 1993;
Mller et al., 1996) and that care must be taken in the equalization and placing
of headphones in order to obtain an immersive 3-D soundscape.
2.2.2 Loudspeaker Reproduction
Loudspeaker reproduction of binaural information is critically dierent from headphone reproduction, as can be seen in Fig. 2.2. The directional characteristics
that are present in binaural signals are exposed to crosstalk when loudspeaker reproduction is used. Crosstalk occurs because the sound from the left loudspeaker
is heard at both left and right ears, and vice versa. The theory of crosstalk
canceling for binaural reproduction was presented in the 1960's by Bauer (1961)
and practically implemented by Schroeder and Atal (1963), and has been later
investigated intensively by the research community and industry. A taxonomy of
crosstalk canceling technologies is given in (Gardner, 1998a).

2.2

31

Methods for 3-D Sound Reproduction

xm

Hl

Hr

yl

xr

xl

yr

Hc

Hc
Hi

Hi

yl

yr

Figure 2.2: Signal owgraph of a) binaural and b) crosstalk canceled binaural

audio processing.
In binaural synthesis for loudspeaker listening (see 2.2b), the crosstalk canceled binaural signals x^l (n) and x^r (n) are applied to the speakers to yield the
desired outputs yl (n) and yr (n). The direction-dependent loudspeaker-to-ear
transfer functions Hi(z) and Hc(z) (we consider here only the symmetrical listening position4) have to be taken into account in order to obtain a similar eect
as in headphone listening. The ltering can here be understood as a cascaded
process, in which HRTF lters are designed and implemented separately from
the crosstalk canceling lters. Another alternative is to combine these processes
and design virtual speaker lters by using, e.g., shuer structures (Cooper and
Bauck, 1989). Shuer structures and other binaural crosstalk canceling systems
will be studied in more detail in Chapter 6. Crosstalk canceled binaural systems
have two main limitations: 1) critical listening position, and 2) critical listening
room conditions. The full spatial information can be retained only in anechoic
chambers or listening rooms and the \sweet spot" for listening is very limited.
On the other hand, at its best crosstalk canceled binaural reproduction is capable
of auralizing a sound eld in a very convincing and authentic way.
2.2.3 Multichannel Techniques
A natural choice to create a two- or three-dimensional auditory space is to use
multiple loudspeakers in the reproduction. With this concept the problems in
retaining spatial auditory information are reduced to placement of the N loudspeakers and panning of the audio signals according to the direction (Gerzon,
1992). The problem of multiple loudspeaker reproduction and implementation
of panning rules can be formulated in a form of vector base amplitude panning
(VBAP) (Pulkki, 1997) or by using decoded 3-D periphonic systems such as
4

The subscript i means ipsilateral (same side), and c means contralateral (opposite side).

32

2. Virtual Acoustics

Ambisonics (Gerzon, 1973; Malham and Myatt, 1995). The VBAP concept introduced by Pulkki (1997) gives the possibility of using an arbitrary loudspeaker
placement for three-dimensional amplitude panning. The rapid progress of multichannel surround sound systems for home and theater entertainment during the
past decade has opened wide possibilities also for multiloudspeaker auralization.
Digital multichannel audio systems for movie theaters and home entertainment
oer three-dimensional spatial sound that has been either decoded from twochannel material (such as Dolby ProLogic) or uses discrete multichannel decoding
(such as Dolby Digital). The ISO/MPEG-2 AAC audio coding standard oers
5.1 discrete transparent quality channels at a rate of 320 kb/s (compression rate
1:12) (Brandenburg and Bosi, 1997). Another multichannel compression technique that is already widespread in the market is Dolby Digital which provides
similar compression rates to MPEG-2 AAC (Brandenburg and Bosi, 1997).
A general problem with multichannel (N > 2) sound reproduction especially
in domestic setups is the amount of needed hardware (6 speakers for a discrete 5.1
surround setup) and their placing. On the other hand, intuitively, the listening
area should be larger and the localization eect more stable than with two-channel
binaural loudspeaker systems. Another advantage of multichannel reproduction
over binaural systems is that no listener modeling (use of HRTFs) is necessary,
and thus it is in general computationally less demanding.
2.3

Implementation of Virtual Acoustic Systems

The previous section dealt with rendering free-eld three-dimensional auditory
cues using dierent reproduction systems. However, when an interactive timevarying virtual acoustic processing scheme (see Fig. 2.1) is considered, the reproduction models interact and are preceded by dierent modeling techniques
for source and room acoustics. Implementation issues of virtual acoustic systems
have been studied by the DIVA group (Takala et al., 1996; Savioja et al., 1999),
including major contributions from the author of this work. In this section, an
outline for a geometrical acoustics-based modeling approach is given.
Real-time rendering of virtual acoustics poses several implementational and
computational challenges. The most crucial ones in terms of perception and
immersion are dynamic (time-varying) room impulse processing and interactivity
of the user with the virtual acoustic model. To gain understanding of room
acoustics modeling techniques, Fig. 2.3 shows dierent methods used in many
current systems (Savioja et al., 1999).
Kleiner et al. (1993) have divided the modeling techniques for room acoustics
between computational simulation and scale modeling. Scale models are widely
used by architects and acousticians in the design of concert halls. However,
computers have been used for thirty years to model room acoustics (Krokstad
et al., 1968; Schroeder, 1970, 1973) and computational modeling has become a

2.3

33

Implementation of Virtual Acoustic Systems

MODELING OF ROOM ACOUSTICS

COMPUTATIONAL
MODELING

WAVE-BASED
MODELING

DIFFERENCE
METHODS

ELEMENT
METHODS

RAY-BASED
MODELING

RAYTRACING

ACOUSTIC-SCALE
MODELING

STATISTICAL
MODELING

DIRECT
MODELING

INDIRECT
MODELING

IMAGE-SOURCE
METHOD

Figure 2.3: Dierent methods for room acoustics modeling and auralization
(Savioja et al., 1999).
widely accepted tool in room acoustic design (Kuttru, 1995).
There are three dierent approaches in computational modeling of room
acoustics as illustrated in Fig. 2.3. In ray-based methods the eect of wavelength is neglected and sound is supposed to act like rays (Kuttru, 1991). All
phenomena due to the wave nature, such as diraction and interference, are excluded from the model. This assumption is valid when the wavelength of the
sound is small compared to the dimensions of surfaces and obstacles in the room
and large compared to their roughness. Ray tracing and image-source methods
are the most often used algorithms in ray-based modeling of room acoustics.
Enhancements to geometrical acoustics modeling principles that take into account diusion (Dalenback, 1995; Naylor and Rindel, 1994) and edge diraction
(Svensson et al., 1997) have been proposed in recent years.
More accurate results can in principle be achieved with wave-based methods
which are based on solving the wave equation. An analytical solution for the wave
equation can be found only in rare cases such as rectangular rooms with rigid
walls. Therefore, numerical wave-based methods must be applied. These methods
are, however, computationally very expensive. Element methods, such as the
nite element method (FEM) and boundary element method (BEM), are widely
used in vibration mechanics, and, despite the heavy computational requirements,
also in room acoustics. As an alternative to element methods, nite-dierence
time-domain (FDTD) simulation has been recently found to be attractive for
room acoustics simulation (Botteldooren, 1995; Savioja et al., 1995).
In addition there are also statistical modeling methods, such as statistical energy analysis (SEA) (Lyon and DeJong, 1995). Those methods are mainly used
in prediction of noise levels in coupled systems where sound transmission through

34

2. Virtual Acoustics

Amplitude

1
0.5

Left Channel

0
Late Reverberation (RT60 ~2.0 s)

0.5

Early Reflections (< 80100 ms)

Direct Sound

1
0

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Amplitude

1
0.5

Right Channel

0
0.5
1
0

0.2

0.4

0.6

0.8

1
1.2
Time / s

1.4

1.6

1.8

Figure 2.4: A binaural room impulse response measured in a concert hall. The
room impulse response can be divided into the direct sound, early re ections,
and late reverberation.
structures is an important factor, but they are not suitable for auralization purposes.
2.3.1 Real-Time Geometrical Room Acoustics Approach
Although the geometrical acoustics modeling approach has known limitations, it
is practically the only relevant technique suitable for real-time or near-real-time
rendering. The output from all virtual acoustic simulators is essentially a binaural
room impulse response (BRIR), an example of which is shown in Fig. 2.4. This is
a two-channel response dependent on the positions and orientations of the source
and the listener and on the properties of the room. An interactive auralization
model should therefore produce output which depends on the dynamic properties
of the source, receiver, and environment. In principle, there are two dierent
approaches to achieve this goal. The methods presented in the following are called
direct room impulse response rendering and parametric room impulse response
rendering.

2.3

Implementation of Virtual Acoustic Systems

35

Direct Room Impulse Response Rendering

The direct room impulse response (DRIR) rendering technique is based on a priori stored BRIRs, obtained either from simulations or from measurements5 . The
BRIRs are dened at a grid of listening points. The auralized sound is produced
by convolving the excitation signal with BRIRs of both ears. In interactive movements this convolution kernel is formed by interpolating BRIRs of neighboring
listening points. The spatial (positional) accuracy of the simulation is dependent on the grid density of BRIR approximation points. The direct convolution
can be carried out in the time or frequency domain or by using hybrid methods (Gardner and Martin, 1994; Jot et al., 1995; McGrath, 1996). Nevertheless,
this technique is computationally expensive and has high memory requirements
for BRIR storage. A further setback of this method is the fact that the source,
room, and receiver parameters can not be extracted from the measured BRIR.
This means that changes in any of these features will require a new set of BRIR
measurements or simulations.
Parametric Room Impulse Response Rendering

A more robust way for interactive auralization is to use a parametric room impulse
response (PRIR) rendering method (Savioja et al., 1999). In this technique the
BRIRs in dierent positions in the room are not predetermined. The responses
are formed in real time during interactive simulation. The actual rendering process is carried out in several parts decomposed in the time domain. The rst part
consists of direct sound and early re ections modeling, both of which are timeand place-variant. The latter part of rendering represents the diuse reverberant
eld, which can be treated as a time-invariant digital lter. In practice this means
that the late reverberator can be predetermined, but the direct sound and early
re ections are auralized according to parameters obtained using a real-time room
acoustic model. The parameters for the articial late reverberation algorithm are
derived from measurements or from room acoustic simulations. This modeling
framework is illustrated in Fig. 2.5 (Savioja et al., 1999). An obvious drawback
of this real-time image-source based model is, however, that it lacks diusion
and diraction models. This fact is recognized and can sometimes degrade the
physical accuracy of the simulation signicantly.
The PRIR rendering technique has been further divided into two categories6
(Vaananen and Huopaniemi, 1999; Jot, 1999):
 Physical modeling approach
 Perceptual modeling approach
5

Dummy head recordings which were the early form of binaural reproduction would also
fall into this category.
6 Jean-Marc Jot. Personal communication. CCRMA, Stanford University, Aug. 6, 1998.

36

2. Virtual Acoustics

ROOM GEOMETRY
MATERIAL DATA

REAL-TIME SYNTHESIS
IMAGE-SOURCE
METHOD

NON-REAL-TIME ANALYSIS
DIFFERENCE
METHOD

RAY TRACING

MEASUREMENTS

ACOUSTICAL ATTRIBUTES
OF THE ROOM

DIRECT SOUND AND

EARLY REFLECTIONS

ARTIFICIAL LATE
REVERBERATION

AURALIZATION

Figure 2.5: Computational room acoustic rendering methods used in the DIVA
system. The model is a combination of a real-time image-source method and
articial reverberation which is parametrized according to room acoustical parameters obtained by simulation or measurements (Savioja et al., 1999).
The physical modeling approach, which is the main topic area of this work, aims
at capturing both the macroscopic (reverberation time, room volume, absorption area, etc.) and microscopic (re ections, material absorption, air absorption,
directivity, diraction, etc.) room acoustical features by a geometrical acoustics approach, and diuse late reverberation using statistical means. The goal
in perceptual parametrization, on the other hand, is to nd an orthogonal set
of features using which a normative virtual acoustic rendering algorithm can
be controlled to produce a desired auditory sensation. Therefore the output is
user-controllable, not environment-controllable as in the physical approach. The
perceptual approach uses physical properties for the direct sound, macroscopic
room acoustical features and a static image-source method for early and late early
re ections, and a statistical late reverberation module for late reverberation (Jot
et al., 1995; Jot and Chaigne, 1996; Jot et al., 1998a; Jot, 1999).
This division clearly separates the two main application goals for parametric

2.4

37

DIVA System

room impulse rendering techniques. The rst approach is suitable for accurate
physical simulation of spaces such as concert halls, auditoria, and for auditorily
accurate audiovisual rendering. The latter approach is computationally more
ecient, has more intuitive perceptually based control parameters, and is thus
more suitable to virtual reality applications and game technology.
The physical PRIR rendering method for binaural reproduction (headphones
or a pair of loudspeakers), which is the main area of interest in this work, will
now be investigated more thoroughly. This technique has been used in the DIVA
system (Takala et al., 1996; Savioja et al., 1999)7. In Fig. 2.6, this modeling
framework is shown, which is based on the well-known and established imagesource method. If we denote the monophonic input signal by xm (n) and the
binaural output from the direct sound and early re ections module as y(n) =
[yl (n) yr (n)]T , the system transfer function H(z) = Y(z)=X (z), where X (z) and
Y(z are z-transforms of xm (n) and y(n), can be written as:
H(z) =

N
X
k=0

Tk (z )Fk (z )z

dk ;

(2.1)

where Tk (z) are cascaded lter structures comprising source directivity (studied
in Chapter 3), material re ection, and air absorption ltering (studied in Chapter
4), F(k z) = [Flk (z)Frk (z)]T are the binaural lters for the direct sound and early
re ections (studied in Chapter 5), and the delay term z dk is the travel time delay
of the direct sound or the corresponding early re ection. The direct sound is fed
(and possibly pre-ltered (Jot et al., 1995)) to the late reverberation module R(z),
which produces diuse late reverberant energy according to the given parameter
specications (see, e.g., (Jot et al., 1995; Vaananen et al., 1997; Savioja et al.,
1999) for more details). Furthermore, if listening is carried out using a pair of
loudspeakers, an additional crosstalk canceling lter network C (z) is required.
Methods and structures for crosstalk canceling are presented in Chapter 6.
2.4

DIVA System

In this section, a short overview of the DIVA system is given. The goal in the research of the DIVA group has been to make a real-time environment for immersive
audiovisual processing. The system should integrate the audio signal processing
chain from sound synthesis through room acoustics simulation to spatialized reproduction. This is combined with synchronized animated motion. A practical
application of this project has been a virtual concert performance (Hiipakka et al.,
7

Similar DSP methods have been proposed and used for real-time auralization by many
researchers. See, e.g., (Foster et al., 1991; Begault, 1994; Jot et al., 1995; Shinn-Cunningham
et al., 1997).

38

2. Virtual Acoustics

direct sound and early reflections

late reverberation

binaural output

sound input

...

x ( n)
m

source directivity,
air and material
absorption,
1/r attenuation
directional
filtering
(ITD+HRTF)

-d 0

-d 1

T ( z)

T ( z)

(optional)

z-d N

...

T ( z)

late reverberation
unit, R (z)

...
F ( z)
0

F ( z)
1

...

out(left)

crosstalk
canceling
C (z)

out(right)

F ( z)
N

...
...

y ( n)
r

y ( n)
l

Figure 2.6: Implementation of virtual acoustics using geometrical acoustics modeling. The direct sound, early re ections and late reverberation are processed
separately, enabling a fully parametric implementation.
1997; Lokki et al., 1999) and virtual acoustic (and visual) simulations of concert
halls and other acoustical spaces (Savioja et al., 1999).
In Fig. 2.7 the architecture of the DIVA virtual concert performance system is
presented. There may be two simultaneous users in the system, a conductor and
a listener, who can both interact with the system. The conductor wears a tail
coat with magnetic sensors for tracking. With hand movements the conductor
controls the orchestra, which may contain both real and virtual musicians. The
aim of the conductor gesture analysis is to synchronize electronic music with
human movements. In the graphical user interface (GUI) of the DIVA system,
animated human models are placed on stage to play music from MIDI les. The
virtual musicians play their instruments at tempo and loudness shown by the
conductor. Instrument ngerings are found from predened grip tables. All the
joint motions of human hand are found by inverse kinematics calculations, and
they are synchronized to perform exactly each note on an animated instrument
model. A complete discussion of computer animation in the DIVA system is
presented in (Hanninen, 1999).
At the same time, a listener may freely y around in the concert hall. The GUI
sends the listener position data to the auralization unit which renders the sound
material provided by physical models and a MIDI synthesizer. The auralized
output is reproduced either through headphones or loudspeakers.
In order to perform real-time auralization according to Fig. 2.6, the following
information is needed before the startup of the processing (Savioja et al., 1999):
 Geometry of the room
 Materials of the room surfaces

2.4

39

DIVA System

Listener

Conductor

Listener
movements
Ascension
MotionStar

Mot

ion

data

tion

iza

n
hro

nc

Sy

Listener
position data

Midi

rol
cont

CCCC
CCCC

MIDI
Synthesizer
for drums

Display
Animation &
visualization
User interface
Binaural reproduction
either with headphones

Instrument audio
(ADAT, Nx8 channels)
Optional ext.
audio input

Physical modeling
Guitar synthesis
Double bass synthesis
Flute synthesis
Conductor gesture
analysis

or with
loudspeakers

Image source
calculation
Auralization
direct sound and early reflections
binaural processing (HRTF)
diffuse late reverberation

Figure 2.7: The DIVA virtual concert performance system architecture (Savioja
et al., 1999).
 Location and orientation of each sound source
 Location and orientation of the listener

Using this information, the early re ections can be dynamically computed using
the image-source method as shown in Fig. 2.6. The parameters concerning each
visible image source that are sent to the auralization module consist of (Takala
et al., 1996; Savioja et al., 1999):
 Distance from the listener
 Azimuth and elevation angles with respect to the listener
 Source orientation with respect to the listener
 Set of lter coecients which describe the material properties in re ections
This parametrization used in the DIVA system has also formed the basis for the
author's contributions to MPEG-4 standardization of three-dimensional audio in
virtual environments. The concepts of scene description in multimedia standards
will be brie y discussed in the next section.

40
2.5

2. Virtual Acoustics

Virtual Acoustics in Object-Based

Multimedia

One prominent future application area of virtual acoustics is virtual three-dimensional environments. Currently, there are several specications that can be used
to create and render audiovisual scenes in multimedia applications that require efcient and parametric coding of objects. Such applications are, for example, those
used across a computer network or other bandwidth restricted delivery channels,
and that need to have interaction and real-time rendering capability. New multimedia standards such as MPEG-4 (ISO/IEC, 1999) and VRML97 (ISO/IEC,
1997) (also commonly known as VRML2.0) specify hierarchical scene description
languages that can be used to build up 3-D audiovisual scenes where the user
can interact with objects. Both VRML and MPEG-4 scene description rely on
a scene graph paradigm to describe the organization of audiovisual material. A
scene graph represents content as a set of hierarchically related nodes. Each node
in the scene graph represents an object (visual objects such as a cube or image,
or a sound object), a property of an object (such as the textural appearance of a
face of a cube, or the acoustical properties of the material), or a transformation
of a part of the scene (such as a rotation or scaling operation). By connecting
multiple nodes together, object-based hierarchies are formed.
The VRML97 standard allows the inclusion of short audio clips to the scenes
in a specied spatial location to improve the immersiveness and interaction between the user and the virtual environment audio. The MPEG-4 audio standard
provides a powerful toolset for both natural and synthetic audio coding and for
audio post-processing and three-dimensional rendering. A binary format scene
description language (binary format for scenes, or BIFS) dened in the systems
level of the MPEG-4 standard enables simultaneous real-time decoding of video
and audio objects and presentation of these elementary stream objects based on
the scene description to the end user. This requires more eciency from the
scene decoding, but allows inclusion of more natural audiovisual objects where
the sound is streamed from the server and can, for example, be real-time encoded
audio, or synthetic speech or audio presented in a parametric form.
The MPEG-4 version 2 standard will provide parameters for virtual acoustic rendering that enable room-dependent eects (reverberation time, material
re ections, air absorption, Doppler eect) and source-dependent features (directivity, attenuation) (Scheirer et al., 1999; Vaananen and Huopaniemi, 1999). In
other words, the desire is to visualize and auralize an MPEG-4 scene with the
given parameters. Listener modeling (HRTF reproduction, crosstalk canceling,
multichannel panning), however, will not be dened (and therefore normative)
in the standard. Spatial presentation of sound in scene description languages
has been studied in greater detail in (Vaananen and Huopaniemi, 1999; Scheirer
et al., 1999). The author of the present work has been a major contributor to

2.6

Conclusions

41

virtual acoustic parametrization of MPEG-4 version 2. However, these issues will

not be dealt with further in this thesis.
2.6

Conclusions

In this chapter, a framework for virtual acoustic modeling was outlined. The
virtual acoustic concept covers source, medium, and listener modeling. The geometrical room acoustics modeling paradigm, although not physically correct especially at lower frequencies, allows for using a parametric room impulse rendering
technique. This time-domain hybrid impulse response reconstruction divided the
modeling into three parts: direct sound, early re ections, and late reverberation
rendering. A structure for dynamic interactive simulation was presented, which
is capable of high-quality virtual acoustic rendering.
In the next chapter, topics in real-time modeling of sound sources in virtual
acoustic systems are covered.

42

2. Virtual Acoustics

Chapter 3
Sound Source Modeling
In this chapter, topics in modeling of sound sources in relation to real-time rendering are studied. The role of sound sources in virtual acoustic environments
is discussed, and ecient sound source directivity rendering techniques are presented. A novel method for modeling sound radiation from the mouth using the
reciprocity method is introduced.
Sound source modeling in the virtual acoustic concept refers to attaching
sound to an environment and giving it properties such as directivity. The simplest approach has traditionally been to use an anechoic audio recording or a
synthetically produced sound as an input to the auralization system (Kleiner et
al., 1993). In an immersive virtual reality system, however, a more general and
thorough modeling approach is desired that takes into account directional characteristics of sources in an eective way. General qualitative requirements for audio
source signals in a virtual acoustic system (used for physical modeling) are:
 Each sound source signal should be \dry," not containing any reverberant
or directional properties, unless it is explicitly desired in the simulation
(e.g., simulating stereophonic listening in a listening room).
 The audio source inputs are normally treated as point sources in the room
acoustical calculation models (image-source method, ray-tracing method,
discussed in Chapter 2). This requires the elementary source signals to be
monophonic. A stereophonic signal emanating from loudspeakers can thus
be modeled as two point sources.
 The quality (signal-to-noise ratio, quantization, sampling rate) of the sound
source signal should be adequately high so as not to cause undesired eects
in auralization.
From the coding and reproduction point of view, audio source signals may be
divided into two categories:
 Natural Audio
43

44

3. Sound Source Modeling

 Synthetic Audio

The concept of natural audio refers to sound signals that have been coded from
an existing waveform. This category includes all forms of digital audio, raw
or bitrate-reduced, and thus the basic requirements for usage in virtual reality
sound are those stated in the previous section. By using various bitrate reduction
techniques, relatively ecient data transfer rates for high-quality digital audio
may be achieved (Brandenburg and Bosi, 1997).
Purely synthetic audio may be explained as sound signal denition and creation without the aid of an a priori coded sound waveform. Exceptions are sound
synthesis techniques that involve wavetable data and sampling, which both use
a coded waveform of some kind. These methods can be regarded as hybrid natural/synthetic sound generation methods. On the other hand, many sound synthesis techniques such as FM (frequency modulation), granular synthesis, additive
synthesis and physical modeling are purely parametric and synthetic1 .
Text-to-speech synthesis (TTS) techniques use a set of parameters that describe the input (often pure text) as phonemes and as prosodic information.
This also allows for very low bitrate transmission, because the speech output
is rendered in the end-user terminal. Furthermore, it is possible to attach face
animation parameters to TTS systems to allow for creation of \talking heads" in
advanced virtual reality modeling systems such as MPEG-4 (ISO/IEC, 1999).
The advantage of using synthetic audio as sound sources in virtual acoustic
environments is the achieved reduction in the amount of data transferred when
compared to natural digitally coded audio. This results, however, in added computation and complexity in the audio rendering. Parametric control provided by
synthetic audio is also a remarkable advantage, allowing for exible processing
and manipulation of the sound (Vercoe et al., 1998).
An attractive sound analysis and synthesis method, physical modeling, has
become a popular technique among researchers and manufacturers of synthesizers
in the past few years (Smith, 1997, 1998; Tolonen et al., 1998). Physical model
based sound synthesis methods are particularly suitable to virtual acoustics simulation due to many reasons. First, one of the very aims in virtual acoustics is to
create models for the source, the room, and the listener that are based on their
physical properties and can be controlled in a way that resembles their physical
behavior. Second, source properties such as directivity may be incorporated into
the physical modeling synthesizers (Huopaniemi et al., 1994; Karjalainen et al.,
1995).
1

See (Smith, 1991; Roads, 1995; Tolonen et al., 1998) for a taxonomy of modern sound
synthesis techniques.

3.1

Properties of Sound Sources

3.1

45

Properties of Sound Sources

Modeling the radiation properties and directivity of sound sources is an important

(yet often forgotten) topic in virtual reality systems. During the course of this
study, the directional properties of musical instruments (Huopaniemi et al., 1994;
Karjalainen et al., 1995; Savioja et al., 1999) and the human head (Huopaniemi et
al., 1999a) were investigated, and real-time algorithms for frequency-dependent
directivity ltering for a point-source approximation were implemented. In general, mathematical modeling of the directivity characteristics of sound sources
(loudspeakers, musical instruments) can be a complex and time-consuming task.
Therefore, in many cases measurements of directivity are used to obtain numerical data for simulation purposes (Meyer, 1978).
3.1.1 On the Directivity of Musical Instruments
String instruments exhibit complex sound radiation patterns due to various reasons. The resonant mode frequencies of the instrument body account for most
of the sound radiation (Fletcher and Rossing, 1991). Each mode frequency of
the body has a directivity pattern such as monopole, dipole, quadrupole, or their
combination. The sound radiated from the vibrating strings, however, is weak
and can be neglected in the simulation. In wind instruments, the radiation properties are dominated by outcoming sound from various parts of the instrument
(the nger holes or the bell). For example, in the case of the ute, the directivity
is caused by radiation from the embouchure hole (the breathier noisy sound) and
the toneholes (the harmonic content) as discussed in (Karjalainen et al., 1995).
Another noticeable factor in the modeling of directivity is the directional
masking and re ections caused by the player of the instrument. Masking plays
an important role in virtual environments where the listener and sound sources
are freely moving in a space. A more detailed investigation of musical instrument
directivity properties can be found, e.g., in (Meyer, 1978). Comparisons between
instrument directivity with and without the player have been published by Cook
and Trueman (1998).
3.1.2 Source Directivity Models
For real-time simulation purposes, it is necessary to derive simplied source directivity models that are ecient from the signal processing point of view and as
good as possible from the perceptual point of view. Dierent strategies for implementing directivity in physical models of musical instruments have been proposed
by the author in several studies (Huopaniemi et al., 1994; Karjalainen et al., 1995)
in relation to sound sources in virtual environments. Of these directivity modeling methods, there are two strategies that are attractive to virtual reality audio
source simulation in general: 1) directional ltering and 2) a set of elementary

46

3. Sound Source Modeling

D1(z)

Sound
source

y^ (n)

D2(z)
DM (z)

y1 (n)

Sound
source

y2 (n)

y1 (n)
y2 (n)
yM(n)

yM(n)

Figure 3.1: Two methods for incorporating directivity into sound source models:
(a) directional ltering, (b) set of elementary sources (Karjalainen et al., 1995;
Savioja et al., 1999).
sources. In Fig. 3.1, these two methods are illustrated. To obtain empirical data,
measurements of source directivity in an anechoic chamber were conducted for
two musical instruments: the acoustic guitar and the trumpet (Huopaniemi et
al., 1994; Karjalainen et al., 1995; Valimaki et al., 1996)2. In both cases, two
identical microphones were placed at a 1 meter distance from the source and the
player, one being in the reference direction (normally 0 azimuth and elevation)
and the other being in the measured direction. An impulsive excitation was used
and the responses H (z) and Hm (z) were registered simultaneously. For modeling purposes, the directivity relative to main-axis radiation was examined, that
is, the deconvolution of the reference and measured magnitude responses was
computed for each direction:
(3.1)
jDm (z)j = jjHHm((zz))jj

In the following, the two main approaches for real-time source directivity approximation are studied.
0

3.1.3 Directional Filtering

The directivity properties of sound sources may be eciently modeled in conjunction with geometrical acoustics methods such as the image-source method. In this
technique, the monophonic sound source output y^(n) is fed to M angle-dependent
digital lters Dm (z) (m = 1; : : : ; M ) representing each modeled output direction from the source (see Fig. 3.1a). Generally, the lowpass characteristic of the
direction lter increases as a function of angle, and low-order lters have been
found to suce for real-time models.
As an example, a set of rst-order IIR lters modeling the directivity characteristics of the trumpet (and the player) is represented in Fig. 3.2. The directivity
2

More extensive instrument radiation databases have been measured and published by
Cook and Trueman (1998) and Semidor and Couthon (1998).

3.1

47

Properties of Sound Sources

Magnitude (dB)

0
-5
-10
-15
-20
0

0
50

100
150
Azimuth Angle ()

10
Frequency (kHz)

Figure 3.2: Modeling the directivity characteristics of the trumpet with rst-order
IIR lters. The azimuth interval of modeled responses is 22:5 (Karjalainen et
al., 1995).
functions were obtained using Eq. (3.1). A rst-order IIR lter was tted to each
result. The lters were designed with a minimum-phase constraint. In our measurements, we analyzed directivity only on the horizontal plane (0 elevation).

3.1.4 Set of Elementary Sources

The radiation and directivity patterns of sound sources can also be distributed,
and a point-source approximation may not be adequate for representation. Such
examples could be, e.g., a line source. So far only a point source type has
been considered and more complex sources have been modeled as multiple point
sources. The radiation pattern of a musical instrument may thus be approximated by a small number of elementary sources. These sources are incorporated
in, e.g., the physical model and each of them produces an output signal as shown
in Fig. 3.1b. (Huopaniemi et al., 1994; Karjalainen et al., 1995). This approach
is particularly well suited to utes, where there are inherently two point sources
of sound radiation, the embouchure hole and the rst open tone hole.

48
3.2

3. Sound Source Modeling

Sound Radiation from the Mouth

In this section, a case study on modeling sound radiation from the human mouth
is presented. Mathematical and empirical methods are discussed, and a novel
technique based on the principle of reciprocity is introduced (Huopaniemi et al.,
1999a).
The free-eld radiated sound spectrum from the mouth varies as a function
of distance and angle. The distance attenuation is known to be caused by the
transmitting medium, air, whereas the directional attenuation is caused by the
human head and torso. The angular dependence in uences the perceived speech
spectrum and intelligibility. In recent years, incorporating directional properties
of sound sources have become attractive in virtual reality and multimedia systems
(Savioja et al., 1999). An example of such specication is the MPEG-4 Systems
scene description language (BIFS) (Koenen, 1999; Swaminathan et al., 1999)
and its advanced audio rendering capabilities (Scheirer et al., 1999). MPEG-4
features human face and body animation, text-to-speech synthesis technology
as well as advanced audiovisual scene description capabilities, thus interactive
virtual reality modeling of human speakers is becoming more and more attractive
with a wide range of application areas.
Directional ltering caused by the human head and torso has been investigated by Flanagan (1960, 1972). Meyer (1978), Marshall and Meyer (1985), did
further work on practical analysis of the directivity of singers. Sugiyama and
Irii (1991) extended a spherical head directivity model to an analytical model
of a prolate spheroid in an innite bae, which was claimed to provide a closer
approximation to the radiation directivity properties of human or dummy head
mouth. Generally it is possible to take both an analytical and empirical approach to the problem of obtaining directivity data. Simple analytical models
are relatively easy to compute and produce moderately good results when compared to empirical data (Flanagan, 1960). Empirical directivity measurements,
on the other hand, can be fairly easily accomplished using a dummy head, but
measurements using human test subjects are very dicult due to, for example,
problems in transducer placement at the mouth opening (if real speech is not
used as input).
An interesting observation can be made here concerning spherical head directivity and its relation to spherical head HRTFs (head-related transfer functions).
Many authors have introduced simplied HRTFs based on approximating pressure distribution around a rigid sphere the size of a human head (Rayleigh, 1904;
Cooper, 1982; Rabinowitz et al., 1993; Blauert, 1997; Duda and Martens, 1998).
But, based on the principle of reciprocity, we are able to use the same data also
for approximating directivity caused by the human head by exchanging the source
and receiver positions (Morse and Ingard, 1968, pp. 342-343). This statement is
valid if a simple spherical point source approximation is used. The main novelty
in this study (Huopaniemi et al., 1999a) has been the application of the reci-

3.2

Sound Radiation from the Mouth

49

procity method to practical directivity measurements; by exchanging the source

and receiver positions a very straightforward measurement can be carried out.
3.2.1 Analytical Formulation of Head Directivity
Analytically, the spherical head directivity can be estimated by examining the
diraction around the rigid head and exchanging the source and receiver positions by introducing the principle of reciprocity. The boundary conditions to be
fullled are (Blauert, 1997):
 At distance r >> a wave propagation is assumed to be planar (r is the
distance from the source to the center of the sphere, and a is the radius of
the sphere).
 At the surface of the sphere the normal component of the volume velocity
is zero.
A formula for range-dependent diraction coecient calculation has been derived
in (Rabinowitz et al., 1993) and (Duda and Martens, 1998). The pressure ps along
the surface of a sphere of radius a is given by Duda and Martens (1998, Eqs. 2-3):
i cS
(3.2)
ps(r; a; f; ; t) = 0 2! e i2ft ;
4a
where 0 is the density of air, c is the speed of sound, f is frequency,  is the
azimuth angle, and S! is the magnitude of ow of an ideal point source. The
variable is the series expansion
1
X
kr)
(3.3)
= (2m + 1) Pm (cos ) hh0m((ka
) ; r < a;
m
m=0
where Pm is the Legendre polynomial of order m, k = 2f=c is wave number,
hm (kr) is the spherical Hankel function of order m in kr, and h0m (ka) is the
derivative of spherical Hankel function of order m with respect to ka. By exchanging the source and receiver positions, the directivity relative to main-axis
radiation ( = 0) can thus be calculated as:
ps (r; a; f; ; t)
d =
:
(3.4)
ps (r; a; f;  = 0 ; t)
In Fig. 3.3, a three-dimensional plot of the spherical head approximation directivity for azimuth angles 0 180 is shown. The lowpass characteristic behavior
is clearly seen3 . In the next section, the empirical measurement method for head
directivity is discussed, and results for the reciprocity measurement principle are
provided.
The \lobing" or \bright spot" eect at 180 azimuth is clearly visible in Fig. 3.3. This
is caused by superposition of arriving signals in-phase, and is likely to be reduced in real
measurements.
3

50

3. Sound Source Modeling

0
5
10
15
20
25
2

10
30
0

10
50
100
150

10

Figure 3.3: Spherical head approximation directivity at 2 meter distance. X-axis:

Azimuth angle (degrees), Y-axis: Frequency (Hz), Z-axis: Magnitude (dB).
3.2.2 Measurement of Head Directivity
In order to compare and investigate the measurement of radiation directivity
characteristics from a human mouth in relation to analytical methods described
in the previous section, a set of measurements was carried out. The goal was to
compare a direct measurement of directional radiation to the reciprocity method
where the source and receiver positions are exchanged. The measurements were
performed in the large anechoic chamber of the Acoustics Laboratory at the
Helsinki University of Technology. The measurements consisted of two parts, in
both of which a Bruel&Kjaer BK4128 dummy head was placed on an electronic
turntable which was rotated at 10 azimuth angular intervals.
The test signal used in the measurements was a 8192 samples long pseudorandom ( at spectrum random phase) signal with a sampling frequency of 48000
Hz. Averaging of 30 measurements was used to improve the signal-to-noise ratio.
The measurements were carried out using a signal processing and measurement
program QuickSig (Karjalainen, 1990) in the Apple Macintosh platform. The
test signal was generated with a dual channel 16-bit AD/DA card installed in the
computer. (Possible interferences of the sound card were eliminated by using the
other channel as a reference by connecting it directly from output to input.) The
signal was then amplied (Luxman M-120A) and transmitted through a loudspeaker (a 150mm diameter spherical shaped loudspeaker cabinet with a 4-inch

3.2

Sound Radiation from the Mouth

51

full-range Audax AT100M0 element (Riederer, 1998a)). The sound pressure was
picked up by a miniature measurement microphone (Sennheiser KE 4-211-2), and
then amplied with a microphone preamplier and transmitted to the input of
the sound card.
First the radiation from the mouth was studied in the direct measurement.
Figure 3.4 illustrates the measurement principle. The signal was transmitted
using the built-in loudspeaker in the dummy head's mouth (BK4128). A microphone was set at the same height at the distance of 2 meters from the center of
the head. After registering the data and rotating the turntable, the measurement
was repeated at 10 azimuth increments for a full circle.
To test the principle of reciprocity, another set of measurements was required.
Figure 3.5 illustrates this measurement principle. The microphone was now
placed in the dummy head's mouth at two positions: a) about 1.4 centimeters from the lip plane, and b) at the mouth opening entrance. The spherical
loudspeaker (Audax AT100M0) was used and placed at the same height at a
2 meter distance. In this case the sound pressure was registered in the mouth
and the measurements were repeated as earlier. Theoretically the measurement
conditions for the reciprocity method should be equivalent, i.e., the same transducers and measuring positions should be used. However, as Eq. 3.4 illustrates,
the directivity is computed relative to main axis radiation, thus the eects of the
transducers are not in uencing the nal result.
The impulse responses obtained as measurement results were converted from
QuickSig (Karjalainen, 1990) to Matlab (Mathworks, 1994) for analysis. The
measurement results were limited to the range of 100 Hz to 15000 Hz. For both
tests, the angular frequency responses were divided by the reference response
measured at 0 as derived in Eq. 3.4. Figures 3.6{3.7 illustrate the measurement
results for the direct case (dashed line) and the reciprocal case (solid line). Figure
3.6 corresponds to the measurement point 14mm inside the mouth opening, and
Fig. 3.7 refers to the microphone position at the mouth opening. Generally, for
head directivity, it is clear that as the angle of incidence increases, increased
lowpass ltering results (as is well known from earlier studies), despite of the
slight undulation in the responses. The attenuation of high frequencies becomes
especially signicant as the wavelength approaches, and decreases beyond, the
dimensions of the human head. The lobing eect, clearly present in the analytical
model results, is also shown in the empirical measurement results, although not
as prominently.
Comparing the direct measurement results to the responses from the reciprocity measurement, it can be seen that they match very well (within 1-2 dB
accuracy) up to about 5 kHz. After this the placing of the microphone inside the
dummy head's mouth becomes interfering, since these wavelengths approach the
dimensions of the mouth and the microphone. The dierence in this frequency
range can be as much as 10 dB (not shown in the gures). However, because of
the good results with frequencies from 100 Hz to 5500 Hz, which are signicant

52

3. Sound Source Modeling

Figure 3.4: Measurement installation in the anechoic chamber for the direct
method.
Distance to source (the mouth speaker of the dummy head): 2
meters. Azimuth angle. Rotational direction of the BK4128 dummy head
placed on a turntable. Measurement microphone (Sennheiser KE 4-211-2) on
stand (Huopaniemi et al., 1999a).
1.

2.

3.

4.

for sound radiation in normal speech bandwidth, it can be said that measurement
based on the principle of reciprocity has very satisfactory performance. The measurement position in the reciprocity technique did not seem to remarkably aect
the results in the frequency range of interest, below 5 kHz, as was expected.
3.2.3 Modeling and Measurement Result Comparison
In the previous two sections, the analytical and empirical approaches for obtaining head directivity data were discussed. In this section, the aim is to compare
and discuss the two approaches and their validity. A similar plot to Figs. 3.6{3.7
for directivity characteristics of the spherical head model is shown in Fig. 3.8. It
can clearly be seen that the overall macroscopic characteristics such as lowpass
ltering as the angle on incidence increases and the lobing eect at the symmetry
point in the back of the head are similar in both methods. However, it is also
noticeable that the overall directivity contours by frequency band as a function
of azimuth angle dier remarkably. This is clearly due to the fact that a simple
spherical model is not an adequate replica of a real or dummy head. It has been

3.3

53

Discussion and Conclusions

Figure 3.5: Measurement installation in the anechoic chamber during reciprocity

method. Distance to source (spherical loudspeaker): 2 meters. Azimuth
angle. Rotational direction of the BK4128 dummy head. Measurement
microphone (Sennheiser KE 4-211-2) installed in the mouth of the dummy head.
Spherical loudspeaker (Audax AT100M0) on stand (Huopaniemi et al., 1999a).
1.

3.

2.

3.

5.

argued in the literature (Sugiyama and Irii, 1991) that a prolate spheroid resembles the human head shape more accurately and would predict the directivity
cues more reliably.
3.3

Discussion and Conclusions

In this chapter, topics in sound source characterization and directivity modeling

were overviewed, and ecient digital lter simulations of musical instrument and
human head directivity were illustrated. A novel measurement and modeling
technique for sound radiation from the human mouth has been developed by
the author. Results of empirical and analytical verication of the method were
provided.
In general, sound source models in virtual acoustics have been often neglected.
In this work, methods for sound source characterization have been presented that
can be used particularly in virtual reality applications. Due to the simplicity of
many of the models, these methods may not be suitable for reproducing complex
source radiation patterns. For practical applications in the context of this thesis,

54

3. Sound Source Modeling

90
Directivity of BK4128 dummy head, mouth opening mic position
120

60

150

30

18 dB

180

12 dB

6 dB

0 dB

500 Hz
4000 Hz
2000 Hz
250 Hz

210

330

125 Hz

300

240
270

Figure 3.6: Directivity of the BK4128 dummy head using the direct method
(dashed line) and the reciprocity method (solid line). The reciprocal microphone
position is at the mouth transducer position (14 mm inside the mouth opening).
however, the framework and results have proven to be very relevant.
In the next chapter, topics in geometrical room acoustics modeling are discussed. Emphasis is placed on a parametric room impulse response rendering
technique, which is particularly suitable for dynamic interactive auralization.

3.3

55

Discussion and Conclusions

90
Directivity of BK4128 dummy head, mouth transducer mic position
120

60

150

30

18 dB 12 dB

180

6 dB

0 dB

500 Hz
4000 Hz
2000 Hz
210

330

250 Hz
125 Hz

300

240
270

Figure 3.7: Directivity of the BK4128 dummy head using the direct method
(dashed line) and the reciprocity method (solid line). The reciprocal microphone
position is at the mouth opening.
90
Spherical head model directivity
120

60

150

30

18 dB 12 dB

180

6 dB

0 dB

4000 Hz
2000 Hz

500 Hz
210

330

250 Hz
125 Hz

300

240
270

Figure 3.8: Directivity of the analytical spherical head approximation.

56

3. Sound Source Modeling

Chapter 4
Enhanced Geometrical Room
Acoustics Modeling
This chapter deals with added features to a real-time geometrical room acoustics
modeling scheme presented in Chapter 2. Among the modeled features are re ection and air absorption behavior. Material absorption values are used, e.g., for
room impulse response (RIR) prediction by computational methods such as the
ray-tracing and the image-source technique (Kuttru, 1995; Savioja et al., 1999).
If the material data is given as a measured impulse response or transfer function,
the problem is to reduce it for ecient simulation. The distance-dependent air
absorption is also a signicant factor that needs to be taken into account in accurate simulation. The parametric room impulse response rendering technique,
discussed in Chapter 2.3.1, separates the processing of direct sound and early
re ections from late reverberation. In a dynamic virtual acoustic environment,
the role of the rst re ections is very important in adding to the perception of
the space. In the following, a technique where boundary material and air absorption characteristics are modeled by low-order minimum-phase digital lters
is presented. The lter design takes into account the logarithmic distribution of
octave frequencies as well as the non-uniform frequency resolution of the human
ear by using frequency warping or weighting functions. A spectral error measure
is used to estimate the needed lter order. The lter design techniques are particularly useful in ecient room acoustics simulation and real-time implementation
of virtual acoustic environments.
It is assumed that the sound traveling in the air and re ecting from surfaces
behaves linearly and that the system is time invariant. Such a system may be
modeled by any linear and time-invariant (LTI) technique. Digital ltering is an
especially ecient LTI technique since DSP algorithms have been developed for
fast execution. With this approach the problem appears as how to simulate the
boundary re ections and air absorption in an ecient, accurate, and physically
plausible way. If a transfer function|re ection and/or propagation|is given
based on measured impulse response or transfer function data or on an analytical
57

58

4. Enhanced Geometrical Room Acoustics Modeling

model, the lter design problem is simply to search for an optimal match of data
to given lter type and order and to given error criteria. If the data is available in
a magnitude{only form, such as absorption coecients for octave bands, there is
an additional task of nding proper phase characteristics as well as interpolation
of the sparsely given data in an acoustically meaningful way.
There exist many `standard' modeling techniques using digital ltering approach based on AR (all-pole), MA (FIR), or ARMA (pole-zero) modeling. For
example, in Matlab (Mathworks, 1994), such lter design functions as yulewalk,
invfreqz, and cremez, are available. The selection of a method depends on the
available response data and target criteria of the design. In the experiments it
was found that the least mean squares t to sparsely sampled data of absorption
coecients or air absorption1 (Huopaniemi et al., 1997). The magnitude response
is rst converted to minimum-phase data, which is then converted to an IIR lter
of desired order.
4.1

Acoustical Material Filters

Measurement and modeling of acoustic materials is an important part of room

acoustical simulation. The traditional methods for material measurements are the
standing wave tube technique (ISO, 1996) and the reverberation room measurement (ISO, 1985). Methods that need digital signal processing are the intensity
measurement technique and the transfer function method (Fahy, 1995; Chung
and Blaser, 1980). Results may be given in various forms: re ection (impulse)
response, re ection transfer function (re ection \coecient"), impedance or admittance data, or absorption data. In the literature, absorption coecients are
generally given in octave bands from 125 Hz to 4000 Hz, specifying only the
magnitude of re ection.
The problem of modeling the sound wave re ection from acoustic boundary
materials is a complex one. The temporal or spectral behavior of re ected sound
as a function of incident angle, the scattering and diraction phenomena, etc.,
makes it impossible to use numerical models that are accurate in all aspects.
Depending on the application, more or less approximation is needed (Huopaniemi
et al., 1997).
In this work the focus was on DSP-oriented techniques to simulate sound
signal behavior in re ection and propagation. Thus many important issues were
ignored, such as the angle dependency of re ection, which could be included and
approximated by various methods. Also, attention was not paid to the fact that
material data measured in diuse eld or in impedance tube should be treated
dierently.
The most common characterization of acoustic surface materials is based on
absorption coecients, given for octave bands 125 Hz to 4000 Hz. On the con1

For example, the invfreqz function in Matlab was found to work well.

4.1

Acoustical Material Filters

59

trary, the DSP-based measurement methods yield the re ection impulse response
r(t) or the complex-valued re ection transfer function (re ectance) R(j ! ) =
F fr(t)g, where F is the Fourier transform. Since the absorption coecient is
the power ratio of the absorbed and incident powers, the relation between the
absorption coecient (!) and the re ectance R(j !) is given by
(! ) = 1 jR(j ! )j2
(4.1)
p
whereby jR(j !)j = 1 (!) can be used to obtain the absolute value of the
re ectance when absorption coecients are given2 . The relation between the
normalized impedance Z (j !) and the re ectance R(j !) is
Z (j ! ) 1
R(j ! ) =
(4.2)
Z (j ! ) + 1
which can be used to compute R(j !) when the material impedance (or admittance, its inverse value) is given. This equation is valid for a normally incident
plane wave re ecting from an innite planar surface. Based on the equations
above, material data can be converted to R(j !) or r(t) for lter approximation.
The application of absorption material lter design to concert hall auralization is
discussed next. The polygon model of the venue under study, Sigyn Hall (Lahti
and Moller, 1996), located in Turku, Finland, is shown in Fig. 4.1. The absorption coecient data depicted in Table 4.1 (Naylor and Rindel, 1994; Lahti and
Moller, 1996) was taken from the concert hall design. The material names and
corresponding octave-band absorption coecients are shown in Table 4.1. In
Fig. 4.2, tting of low-order lters to absorption data is shown. In the gure, the
magnitude responses of 16 IIR lters designed to match the corresponding target
values are plotted (the target response at Nyquist frequency was approximated
from the 4 kHz octave band absorption coecient). The lters were designed
using a weighted least squares (LS) approximation using ERB scale weighting.
From the plot it can be seen that for most absorption properties a rst-order IIR
lter is an adequate approximation. For some data, a higher-order (order 3) lter
was needed.

A negative value of the square root is possible in theory but practically never happens in
practice.

60

4. Enhanced Geometrical Room Acoustics Modeling

Figure 4.1: Computer model of Sigyn Hall, Turku, Finland (Lahti and Moller,
1996).
125 Hz

0.01
0.08
0.15
0.55
0.11
0.02
0.08
0.1
0.3
0.25
0.15
0.08
0.02
0.15
0.15
0.16

250 Hz

0.01
0.11
0.1

500 Hz

1000

2000

4000

0.01
0.05
0.06

Hz

Hz

Hz

0.02
0.03
0.04

0.02
0.02
0.04

0.02
0.03
0.05

Description

smooth beton
gypsum 13mm
gypsum
2x13mm
0.7
0.65
0.65
0.6
0.6
gypsum 710mm
0.13
0.05
0.03
0.02
0.03
gypsum 9mm
0.02
0.04
0.05
0.05
0.1
cork
0.04
0.03
0.03
0.02
0.02
glass 1k
0.07
0.05
0.03
0.02
0.02
glass 2k
0.15
0.1
0.05
0.03
0.02
glass 610
0.15
0.1
0.09
0.08
0.07
wooden panels22
0.11
0.1
0.07
0.06
0.07
wooden oor
0.2
0.55
0.65
0.5
0.4
wooden wall4
0.02
0.03
0.04
0.04
0.05
linoleum
0.7
0.6
0.6
0.75
0.75
wool 50mm
0.7
0.6
0.6
0.85
0.9
wool 50mmkas
0.24
0.56
0.69
0.81
0.78
audience
Table 4.1: Sigyn Hall absorption coecient data.

4.1

61

Acoustical Material Filters

Material 1. Order: a=1, b=1

Material 2. Order: a=1, b=1

Material 3. Order: a=1, b=1

0
1
2

Material 5. Order: a=1, b=1

0
1
2

Material 9. Order: a=1, b=1

1
2
0

2
Material 11. Order: a=1, b=1

Material 8. Order: a=1, b=1

Material 6. Order: a=1, b=1

Material 7. Order: a=1, b=1

Magnitude (dB)

Magnitude (dB)

Material 4. Order: a=3, b=3

3
4
5
6
7

Material 10. Order: a=1, b=1

Material 12. Order: a=3, b=3

2
1

Material 13. Order: a=1, b=1

Material 14. Order: a=3, b=3

10

Material 15. Order: a=3, b=3

Material 16. Order: a=3, b=3

0

10
10

15
200

1000
Frequency (Hz)

10000

200

1000
Frequency (Hz)

10000

Figure 4.2: First-order and third-order minimum-phase IIR lters designed to

given absorption coecient data. Solid line: octave-band data, dashed line: IIR
lter approximation.

62
4.2

4. Enhanced Geometrical Room Acoustics Modeling

Air Absorption Filters

The eect of air absorption is an important factor in image-source calculations of

large acoustical spaces, such as concert halls where higher-order re ections can
arrive considerably delayed from the direct sound. The absorption of sound in the
transmitting medium (normally air) depends mainly on the distance, temperature
and humidity. There are various factors which participate in absorption of sound
in air (Kuttru, 1991). In a normal environment the most important of those
is the thermal relaxation. The phenomenon is observed as increasing lowpass
ltering as the distance between the listener and the sound source increases.
For real-time simulation purposes, air absorption was approximated in the
following ecient manner. Analytical expressions for attenuation of sound in
air as a function of temperature, humidity and distance have been published by
Bass and Bauer (1972) and discussed in (Kuttru, 1991). Equations for sound
absorption in air have also been standardized (ISO, 1993). These standard formulae were used for calculation, yielding an attenuation a in dB per meter at the
frequency f with variables T (temperature in Kelvin), h (molar concentration of
water vapor), and pa (ambient sound pressure amplitude):
a = 8:686f

"

1:84  10

11

 0:01275e

+ 0:1068e
where

pa
pr

2239:1

3352:0

 1

fro +

T
T0

1=2 #

 2  1
f

frn +

T
T0

 5=2

(4.3)

fro

 2  1 !
f

frn

p
0:02 + h
fro = a 24 + 4:04  104 h
pr
0:391 + h

 1=2
4:170 ( TT0 )
p T
9
+
280
he
frn = a
pr T 0

1=3

!

(4.4)

where fro is the oxygen relaxation frequency, frn is the nitrogen relaxation frequency, T0 is the reference air temperature (T0 = 293:15 K), and pr is the reference
ambient atmospheric pressure (pr = 101:325 kPa).
In the following, transfer functions and lter design examples for air absorption at 20C and 20% humidity (h = 0:4615) for various distances are presented.
The resulting magnitude responses using Eqs. 4.3{4.4 were tted with IIR lters
(two poles: na = 2, one zero: nb = 1) designed using the technique described
previously. Results of modeling for six distances from the source to the receiver
are illustrated in the top plot of Fig. 4.3. As can be seen from the gure, the

4.3

63

Implementation of Extended Image-Source Model

Air absorption, humidity: 20%, nb: 1, na: 2, type: LS IIR

Magnitude (dB)

0
1m

10 m

20 m
30 m

40 m
50 m

8
10
12
3

10

10

Air absorption + 1/r, humidity: 20%, nb: 1, na: 2, type: LS IIR

0
Magnitude (dB)

1m

10
20
10 m

30

20 m
30 m
40 m
50 m

40
50

10
Frequency (Hz)

10

Figure 4.3: Magnitude response of air absorption lters as function of distance

(1m-50m) and frequency. The continuous line represents the ideal response and
the dashed is the lter response. The air humidity is chosen to be 20% and
temperature 20C. Upper plot: air absorption only. Lower plot: air absorption
+ 1/r distance attenuation.
lowpass characteristic is very considerable at larger distances from the source.
This implies that for perception of the direct sound and rst re ections the air
absorption is a contributing factor.
The distance attenuation (1=r law) also has a profound eect on the gain of
the direct sound and the early re ections. When an 1=r gain is added to the lter
simulations of Fig. 4.3 (top plot), we result in transfer functions depicted in the
lower plot of Fig. 4.3.
4.3

Implementation of Extended Image-Source

Model

The real-time implementation of the extended image-source model follows the

principles depicted in Fig. 2.6 (on page 38). An exploded view of lter structures
Tk (z ) and Fk (z ) = [Flk (z )Frk (z )]T is shown in Figs. 4.4{4.5.
The raw audio input to be auralized is fed to a delay line (as shown in Fig. 2.6).
The delay line length corresponds to the propagation delay from the sound source
and its image sources to the listener. In the next phase all image sources are
ltered with lter blocks Tk (z), which contain the following lters (cascaded or

64

4. Enhanced Geometrical Room Acoustics Modeling

0-N

( z)

( z)

source directivity

( z)

material absorption

AI

( z)

air absorption

0-N

1-N

0-N

distance attenuation

Figure 4.4: Detailed view of the real-time extended image-source modeling algorithm shown in Fig. 2.6 (page 38). The lter structure Tk (z) comprises source
directivity lters Dk (z), material absorption lters Mk (z), air absorption lter
AIk (z ), and a gain factor for the 1=r attenuation law (Savioja et al., 1999).
integrated, depending on system implementation):
 Source directivity lter
 Surface material lters, which represent the ltering occurring at corresponding re ections (not applied to the direct sound).
 The 1=r-law distance attenuation
 Air absorption lter
The signals produced by Tk (z) are then ltered with listener model headrelated transfer function (HRTF) lters Fk (z), which perform binaural spatialization. The recursive late reverberation module R(z) (see Fig. 2.6) is summed
to the outputs from the direct sound and early re ection module.
4.4

Discussion and Conclusions

In this chapter, topics in real-time geometrical room acoustics modeling were

discussed. Absorption due to re ections from boundary materials and the transmitting medium was reviewed. A modeling technique for material and air absorption using low-order digital lters was presented. Implementation aspects of
an extended image-source model that takes into account directivity, and material
and air absorption eects were summarized.
It should be noted that the image-source method has limitations in the capability to accurately model diusion and diraction. On the other hand, it is

4.4

65

Discussion and Conclusions

ITD

( z)

ITD

0-N,L

0-N

( z)

( z)

0-N,R

interaural time
difference

0-N,L

( z)

0-N,R

( z)

minimum-phase
HRTFs

Figure 4.5: Detailed view of the real-time extended image-source modeling algorithm shown in Fig. 2.6 (page 38). The lter structure Fk (z) implements the
interaural time delays IT Dk and the minimum-phase HRTFs Hk (z) (Savioja et
al., 1999).
a practical choice for implementation of the direct sound and early re ections
in a parametric room impulse rendering scheme. The perceptual importance of
improved modeling of re ections and air absorption was not veried in subjective
listening experiments. It is expected that the eect of these features is not significant in static reproduction, but in dynamic simulation, where both the listener
and source can move in the virtual acoustic space, correct implementation of the
features is important.
In the next chapter, topics in binaural modeling are covered. Application and
approximation methods of head-related transfer functions (HRTFs) are explored.

66

4. Enhanced Geometrical Room Acoustics Modeling

Chapter 5
Binaural Modeling and
Reproduction
The accuracy needed in binaural lter design and synthesis is a current issue
in multimedia and virtual environment research. More and more eort is being
laid on computationally ecient binaural synthesis models for headphone and
loudspeaker reproduction (Huopaniemi and Smith, 1999; Jot et al., 1998b).
In this chapter, the basic theory for binaural synthesis is presented rst. Manipulation of modeled or measured HRTFs to suitable form by proper equalization and smoothing techniques is reviewed. Digital lter design aspects of HRTF
modeling are reviewed extensively, and novel methods are proposed that take into
account the non-uniform frequency resolution of the human ear. New objective
and subjective methods of binaural lter design quality estimation are presented,
based on binaural auditory modeling and listening experiments.
In the literature, several review papers on 3-D sound can be found that give
a detailed description and discussion on topics related to this chapter. Such
examples are articles by Mller (1992); Kleiner et al. (1993); Kendall (1995) and
books by Blauert (1997); Begault (1994); Gilkey and Anderson (1997).
5.1

Properties and Modeling of HRTFs

HRTFs are the output of a linear and time-invariant (LTI) system, that is, the
diraction and re ections of the human head, the outer ear, and the torso, in
response to a free-eld stimulus registered at a point in the ear canal. Thus, in
theory, the head-related impulse responses (HRIRs), when properly processed,
can directly be represented as non-recursive nite-impulse response (FIR) lters.
There are often computational constraints, however, that call for HRTF approximation. This can be carried out using conventional digital lter design techniques,
but it is necessary to note that the lter design problem is not a straightforward
one. It should be possible to design arbitrary-shaped mixed-phase lters that
67

68

5. Binaural Modeling and Reproduction

meet the set criteria both in both amplitude and phase response. The main
questions of interest that the lter design expert is faced with are at this point:
What is important in HRTF modeling? Are there constraints in the amplitude
and phase response, and if so, how are they distributed over the frequency range
of hearing? In spatial hearing, it is well known that low-frequency (below approximately 1.5 kHz) interaural time delay (ITD) cues have a dominant role in
localization, whereas the interaural level dierence (ILD) is the major localization cue at higher frequencies. Furthermore, mid- to high-frequency spectral cues
enable elevation detection and out-of-head localization, especially in the median
plane (where interaural cues vanish) and on the cone of confusion (where interaural cues are ambiguous). It is, however, not clear what the salience of these
cues in localization is, and how accurately they should be modeled in binaural
synthesis.
The process of binaural synthesis (see Fig. 2.2 in Section 2.2) can be formulated in terms of an LTI system. If the monophonic input signal is denoted by
xm (n), the resulting binaural signals are obtained by the following convolution:
=  xm (n)
(5.1)




y
(
n
)
h
(
n
)
= yrl (n) ; = hrl (n)
where is a column vector of binaural signals and is a column vector of
HRIRs used for binaural synthesis. Going from discrete time-domain to a discrete
frequency-domain representation, Eq. 5.1 can be rewritten as:
y

=
=

Xm (ej! )

Yl (ej! ) ; H = Hl (ej! )
Yr (ej! )
Hr (ej! )

(5.2)

Next, any single HRTF can be noted by H (ej! ) (for convenience from now
on the term HRTF refers both to a time-domain HRIR and a frequency-domain
HRTF). The frequency response of an HRTF can be decomposed into amplitude
and phase responses:


j!
H (ej! ) = H (ej! ) ej argfH (e )g
(5.3)
Any LTI system function can be further represented as a cascade of a minimumphase and an allpass-phase system (Oppenheim and Schafer, 1989):


j!
j!
H (ej! ) = H (ej! ) ej (argfHap (e )g+argfHmp (e )g)
(5.4)

5.1

69

Properties and Modeling of HRTFs

DATA

Audio
Input

USER INPUT
HRTF Database
= [0:10:350]
= [90:10:90]

ITD
Table

Azimuth
Elevation

OUTPUT
Coefficient
Interpolation

DLl

hl,i (n,,)

DLr

hr,i (n,,)

Crosstalk
Canceling

Loudspeaker
Listening

Headphone
Listening

Figure 5.1: General implementation framework of dynamic 3-D sound using nonrecursive lters. HRTF modeling is carried out using pure delays to represent
the ITD (DLl and DLr ), and minimum-phase reconstructed interpolated HRIRs
(hl;i(n) and hr;i(n)).
In Fig. 5.1, a schematic for 3-D sound processing for a single source using
non-recursive ltering with dynamic interpolation is presented. The user gives
desired azimuth and elevation angles  and , after which the corresponding lter
coecients are calculated by fetching the needed ITD and HRTF (HRIR) values
from a database using table lookup and coecient interpolation. Delay lines
DLl and DLr correspond to the ITD values given for the desired azimuth and
elevation angles, and hl;i(n; ; ) and hr;i(n; ; ) are the interpolated minimumphase HRIRs.
Minimum-Phase Reconstruction

An attractive property of head-related transfer functions is that they are nearly

minimum phase (Mehrgardt and Mellert, 1977; Kistler and Wightman, 1992).
Independent research studies have shown that minimum-phase reconstruction
of HRTFs does not have any perceptual consequences (Kistler and Wightman,
1992; Kulkarni et al., 1995, 1999). A minimum-phase system refers to a transfer
function in which all the zeros (and poles) lie inside the unit circle. It also has the
least energy delay of all realizable systems (with the same magnitude response),
thus suggesting benets in lter design and in interpolation of coecients (in
time-varying implementations). A minimum-phase reconstruction of any system
function can be conveniently carried out based on the property that the real and
imaginary parts of H (ej! ) are related by the Hilbert transform (Oppenheim
and


Schafer, 1989). Therefore, a minimum-phase reconstructed version Hmp(ej! ) of

70

5. Binaural Modeling and Reproduction

a system response jH (ej! )j can be found by windowing the real cepstrum in the
following way (see Oppenheim and Schafer (1989, p. 784)). Let us denote the
cepstral window function yielding a minimum-phase reconstruction by wmp(n),
where
8
>
<

1; n = 1 or n = N=2 + 1;
wmp(n) = 2; n = 2; : : : ; N=2;
(5.5)
>
:
0; n = N=2 + 2; : : : ; N:
for even N . The cepstral windowing is accomplished by rst taking the inverse
fast Fourier transform (IFFT) of the logarithm of the magnitude frequency response (the denition of cepstrum):



c(n) = F 1 flog jF fh(n)gjg = F 1 log H (ej! )
(5.6)
and windowing the cepstrum by the dened function wmp(n):
c^(n) = c(n)wmp (n)
(5.7)
The minimum-phase impulse response is then reconstructed by the following operation:


hmp(n) = F 1 eFfc^(n)g
(5.8)
The decomposition of HRTFs into minimum-phase and allpass components
has many advantages. With minimum-phase reconstructed HRTFs, it is possible
to separate and estimate the ITD of the lter pair, and insert the delay as a
separate delay line (and/or an allpass lter) in series with one of the lters in
the simulation stage (an example is seen in Fig. 5.1). The same delay line may
also be used for implementing a re ection path delay in the image-source method
(Takala et al., 1996; Savioja et al., 1999). The frequency-independent ITD can
be calculated using the methods presented in the following subsection. It should
be noted, however, that the delay error error due to rounding of the ITD to
the nearest unit-delay multiple may cause artifacts in the processing stage. The
maximum angular error caused by ITD rounding can be estimated using Eq. 5.11
for and solving for :

 c
error
(5.9)
error = arcsin
2a
(the maximum ITD error is thus error = 1=fs). For example, at fs = 48 kHz,
the maximum angular error error = 2:3, but at fs = 32 kHz, error increases
to 3:5, when a = 0:0875 m and c = 340 m/s. According to Blauert (1997), the
localization blur (localization error) of an average listener in frontal directions

5.1

Properties and Modeling of HRTFs

71

is approximately 3:6. This would indicate that ITD approximation using an

integer number of samples delay is sucient unless a low sampling rate (fs < 32
kHz) is used. If more accurate control is desired or if the sampling frequency is
decreased, the rounding problem can be avoided using fractional delay ltering
(see (Laakso et al., 1996) for a comprehensive review on this subject).
An example of time- and frequency-domain representations of HRTFs1 are
shown in Figs. 5.2 and 5.3. The measurements were made on a Cortex MK2
dummy head with a sampling rate fs = 48 kHz. The rst 256 samples of the impulse response were used for both the time-domain and the fast Fourier transform
(FFT) plots. It can be seen that the bulk delay is removed from the beginning
of the impulse responses as a result of minimum-phase reconstruction. Furthermore, some of the ne structure of the impulse responses appears changed. This
is evidence that the original HRIRs cannot be made minimum phase by a simple
time shift to remove the initial time delay2 . The magnitude responses shown in
Fig. 5.3 are identical as expected. In the following sections, a closer look at the
basic spatial hearing cues { ITD, ILD, and spectral information { will be taken.
5.1.1 Interaural Time and Phase Dierence
The dierence in the time of arrival of sound at the two ears is one of the basic cues of spatial hearing, especially at low frequencies. The ear is sensitive
to phase-derived low-frequency ITD cues at frequencies below approximately 1.5
kHz, where the phase of the incident sound is uniquely determined (Blauert,
1997). The low-frequency interaural time dierence cues have been found to
dominate sound localization (Wightman and Kistler, 1992). The limit to accurate
ITD cues is caused by the dimensions of the human head, because at higher frequencies the phase of the incident sound is ambiguous. The frequency-dependent
behavior of ITDs has been discussed and modeled in (Kuhn, 1977, 1987). Computational models for ITDs in the horizontal plane were compared and the resulting
approximations were as follows. For low-frequency phase-derived ITDs (ka  1),
it was shown that
3a
(5.10)
 = sin ; for ka  1
low

By denition an HRTF is a free-eld response at a point in the ear canal divided by the
response in the middle of the head with the head absent (Blauert, 1997). Deconvolution was
performed to obtain the results shown in Figs. 5.2 and 5.3.
2 To draw this conclusion, it is necessary to ensure that time aliasing is insignicant in
the computation of the cepstrum, which will occur if there are any notches in the HRTF too
close to the unit circle. A simple but eective method is to verify that the signal energy at the
midpoint of the cepstral inverse-FFT (IFFT) buer is small compared to the total energy in
the buer (Smith, 1983).

72

5. Binaural Modeling and Reproduction

Impulse response, azim= 40, id=2

Amplitude / original

2
1
Right

0
1

Left

2
0

0.5

1.5

2.5

3.5

4.5

5
3

Amplitude / min.phase

x 10
2
1
Right

0
1

Left

2
0

0.5

1.5

2.5
Time / s

3.5

4.5

5
3

x 10

Figure 5.2: Upper plot: Mixed-phase HRIRs. Lower plot: Minimum-phase

HRIRs.
where  is the azimuth angle and k = 2f=c is the acoustic wave number. For
high-frequency ITD (ka >> 1) the approximation is of the form
(5.11)

= 2a sin ; for ka  1
high

It has been shown in (Kuhn, 1977) that these low- and high-frequency ITD
asymptotes match well to measured data on a dummy head. Thus, it can be
summarized that the theoretical ratio of the low-frequency ITD (below 500 Hz)
to the high-frequency ITD (above 2000 Hz) is approximately 3=2 (Kuhn, 1977).
Below and above the asymptotes the ITD behavior is frequency-independent,
but in the transition band (approximately 500-2000 Hz) the ITD has a decreasing slope. In (Woodworth and Schlosberg, 1954), another ITD approximation
is presented, which is based on a physical model and simplied spherical geometry. This model is a frequency-independent, high-frequency approximation for
predicting the ITD. The method of calculation is presented in Fig. 5.4. This
approach yields for parallel sound incidence (far-eld case):

= a (sin + )
(5.12)
group

5.1

73

Properties and Modeling of HRTFs

Magnitude response, azim= 40, id=2

25
20
15
10
Magnitude / dB

Right

5
0
5

Left

10

Original R
Original L
Min.phase R
Min.phase L

15
20
25

10

10
Frequency / Hz

Figure 5.3: Mixed-phase and minimum-phase HRTFs (note that the mixed-phase
and minimum-phase magnitude responses are equal).
Left

A'

Right

a
asin
a

Figure 5.4: ITD approximation based on distance dierence of a sound arriving

at the two ears.

74

5. Binaural Modeling and Reproduction

Estimating the ITD from Measured HRTFs

Apart from theoretical models for ITD, there is often a need to estimate the ITD
from actual measurements of HRTFs. Here a distinction should be made between
methods that are purely a) for estimating the ITD between two HRTFs and b) for
estimating the ITD for a digital lter realization of HRTFs (the decomposition
to minimum-phase and allpass sections). Three techniques are generally used for
derivation of the ITD:
 Computation of the exact interaural phase dierence (IPD) and derivation
of a frequency-independent or frequency-dependent ITD approximation
 The cross-correlation method, where the ITD is computed as the oset of
the cross-correlation maxima of the two impulse responses
 The leading-edge method, where the time dierence of the start of the
impulses (above a threshold) is computed. In this method it is assumed
that the HRTFs are minimum-phase or nearly minimum-phase.
Of these approaches, the rst one naturally gives the exact answer based on the
IPD. The latter methods are sometimes used for deriving a frequency-independent
ITD value. If the signals are not minimum-phase, the three above methods may
give dierent results. For ITD approximation to be used in conjunction with a
minimum-phase digital lter realization, the interaural phase characteristics of
the HRTF lters should also be taken into account and compensated for. In the
following, a method will be outlined that uses minimum-phase reconstruction
and linear interaural excess phase approximation (discussed in, e.g., Jot et al.
(1995)). This method has been used throughout the rest of this study for the
approximation of ITDs for HRTF synthesis.
As discussed by Jot et al. (1995) and subsequently by Gardner (1997), the excess interaural phase response of the HRTFs (the term arg fHap(ej! )g in Eq. 5.4)
can be modeled using linear regression over a frequency band. The desired excess
phase response is rst calculated by deconvolving the original HRTF with its
minimum-phase counterparts. In the second step, a straight line is tted to the
interaural excess phase response over a frequency band (e.g., 500-2000 Hz). In
Fig. 5.5, this minimum-phase and linear excess-phase approximation method is
illustrated using the example HRTF pair. It can also be seen in the gure that
the theoretical low- and high-frequency ITD models governed by Eqs. 5.10 and
5.11 t well to measured data.
Comparison of Empirical and Analytical ITDs

When a computational ITD calculation method is compared to ITDs derived

from HRTF measurements averaged from human subjects (Riederer, 1998a), it
can be seen that the Woodworth model given in Eq. 5.12 matches quite well to the

5.1

75

Properties and Modeling of HRTFs

Phase (unwrapped) / rad

IPD, azim= 40, id=2

0

10

20

Original IPD
Min.phase + linear excess approx.

30

10
0

x 10

10
ITD, azim= 40, id=2

ITD / s

0.2
0.4
0.6

Original ITD
Min.phase + linear excess approx.
LF model (01.5 kHz)
HF model (1.520 kHz)

0.8
1

10

10
Frequency / Hz

Figure 5.5: Estimation of the ITD using the minimum-phase and linear excess
phase approximation method (Jot et al., 1995). Upper plot: Original IPD (solid
line) and IPD approximation (dashed line). Lower plot: Original ITD (solid
line), ITD approximation (dashed line), and low-frequency and high-frequency
ITD models (dash-dot lines).
empirical data. The results for the horizontal plane (elevation=0) are illustrated
in Fig. 5.6. The oset of the ITDs around 90 and 270 suggests that the ears are
placed a little asymmetrically, approximately at 100 and 260 azimuth, which is
a commonly accepted fact (Blauert, 1997). For comparison, the results of highfrequency and low-frequency models given in Eqs. 5.10{5.11 are also depicted.
Elevation Dependency of the ITD

Elevation dependency modeling of the ITD has not been explored extensively in
the literature (Larcher and Jot, 1997; Savioja et al., 1999). This has been mainly
due to the lack of an appropriate model. Empirical results have shown, however,
that the ITD is reduced when the elevation angle is increasing, as expected. The
ITD curves have also been found to be approximately symmetrical with respect
to 0 elevation (e.g., 15 elevation ITD curves are of similar nature). The ITD
measurement data has been found to t well to a spherical head based ITD model

76

5. Binaural Modeling and Reproduction

ITD approximations vs. measurements, a=0.0875 m

-4

x 10
8
6
4

ITD (s)

2
0
-2
-4
Frequency-independent model
Kuhn low-freq model
Kuhn high-freq model
Measured ITDs (33 subjects)

-6
-8
0

50

100

150
200
Azimuth angle (deg)

250

300

350

Figure 5.6: Comparison of ITD approximations to measured ITDs of 33 human

subjects at 0 elevation. ITD computation for measured HRTFs was carried out
using the cross-correlation method.
(given in Eq. 5.12). The elevation dependency of the ITD has been taken into
account by adding a scaling term to the basic ITD equation:
a(sin  + )
elev =
(5.13)
2c cos 
where  is the elevation angle. Another approximation method including the
elevation angle has been proposed by (Larcher and Jot, 1997):
elev =

a
2c (asin(cos  sin ) + cos  sin )

(5.14)

A simple cosine dependency of the elevation angle was found to be accurate

enough for simulation purposes. A plot illustrating the elevation dependency of
ITDs for a test subject is shown in Fig. 5.7. The empirical frequency-independent
ITD values were calculated using the cross-correlation method. It can be seen that
a simple cosine model is able to model the ITD behavior in dierent elevations.
An ellipsoidal head model that can be used to predict elevation ITD behavior
has been presented in (Duda et al., 1999).

5.1

77

Properties and Modeling of HRTFs

x 10

0 deg

5
ITD (s)

ITD measured at different elevations for a human subject

30 deg
15 deg
60 deg

90 deg

5
0
x 10

50

100

150
200
250
Azimuth (degrees)
Modeled elevationdependent ITD

0 deg

ITD (s)

300

350

300

350

30 deg

15 deg
60 deg

90 deg

5
0

50

100

150
200
Azimuth (degrees)

250

Figure 5.7: Measured and simulated ITDs at various elevation angles as a function
of azimuth angle.
5.1.2 Interaural Level Dierence and Spectral Cues
The interaural level dierence (ILD) is the second major cue of Lord Rayleigh's
duplex theory. It is mainly due to direction-dependent and frequency-dependent
shadowing caused by the pinnae, the head, and the torso. The role of the ITD
starts to decrease with frequencies above 2 kHz. This is due to the shortening
wavelength of the incident sound and the resulting weaker phase change perception. Instead, the ILD dominates the localization at higher frequencies. The
eective frequency range of the interaural pressure level dierence (ILD) spans
throughout the audible frequency range, although at lower frequencies the role
of the ITD is dominant. Problems in the localization of sound occur mostly in
the median plane, where the ITD and ILD are small or zero, and on the cone
of confusion, where ITD and ILD cues are ambiguous. Although actual human
heads are not spherical and the ears are asymmetrical in shape and placement,
the proposed mechanisms causing localization mismatch are valid.
The behavior of ILD varies considerably as a function of incident angle and
frequency. Source distance starts to eect the ILD below approximately 1m.
Furthermore, since ILD cues are of individual nature, accurate modeling using
mathematical principles is impossible. In the literature, models for the ILD have

78

5. Binaural Modeling and Reproduction

been computed based on Rayleigh's formulas for diraction around a rigid sphere
(discussed in the context of source directivity modeling earlier in Chapter 3).
Interaural cues, ITD and ILD, both have the eect of causing lateral displacement to the position of the auditory event. Trading of interaural cues (also called
time-intensity trading) has been applied to compensate, e.g., for asymmetrical
listening positions. See, e.g., (Blauert, 1997) for a review on this issue.
The pinna cavities have an important role in adding frequency-dependent
directional behavior to the sound eld emitted to the ear canal. The role of
spectral cues especially at high frequencies in median plane localization and elevation distinction is very important. Spectral cues are also dominant in monaural
localization.
5.1.3 Other Cues for Sound Localization
In the previous section, the process of human sound localization was limited only
to static case and without intersensory cues. It has, however, been shown that
human sound localization may be enhanced, even exaggerated, by introducing additional cues to obtain a better performance, for example, in virtual environment
simulation.
Intersensory Cues

The role of visual cues in the perception of spatial sound must not be underestimated. The visual cues often dominate over spatial hearing cues when perceiving
the direction of the incident sound. The eect called ventriloquism is dened
as the spatially biased perception of the auditory stimulus from the same point
as the visual stimulus (Shinn-Cunningham et al., 1997, p. 620). It is very likely
that ambiguous direction perception of sound sources has a great benet of visual
cues. Furthermore, visual cues may be an important factor in distance judgments
of a sound source.
Room Acoustic Cues

Room acoustic phenomena have been found to in uence and enhance localization.
From room acoustical theory it is known that the early re ections add to spatial
impression and perception of the acoustic space (Barron, 1993). In a small room,
early re ections may resolve localization problems in the median plane or on the
cone of confusion.
Head Movements

It is clear that the accuracy of sound localization is increased with the aid of head
movements. Head movements are especially useful in solving ambiguous directional information, particularly in the \cone of confusion" where front-back and

5.1

Properties and Modeling of HRTFs

79

up-down reversals most often occur. The primary cues for distinguishing front
from back seem to be head movement and pinna response. A study conducted by
Wightman and Kistler (1999) shows (and veries earlier results) that user-induced
head or source movements involving rotation provide signicant reduction in the
horizontal errors of directional localization. A similar study presented on free
head movements during sound localization (Thurlow et al., 1967) showed that
rotation movements of the head about the vertical axis (Z) were most commonly
found alone, or in combinations with rotations about the interaural axis (Y) and
the X-axis. Many subjects also showed reversal motions, that is, moving the
head back and forth during sound localization. These results show that listener
movement and head-tracking may be a signicant factor to aid in the localization
of virtual sources in headphone listening.
Super-Auditory Localization

Some applications that take advantage of 3-D sound benet from models of exaggerated spatial hearing cues. In super-auditory localization, both the ITD
and ILD cues can be modied to match an enlarged human head (Rabinowitz
et al., 1993; Shinn-Cunningham et al., 1997). This results in exceeded localization performance that may be desired in virtual auditory displays designed for
teleoperation and virtual environments.
5.1.4 Distance-Dependency of the HRTF
In this section the characterization of distance in HRTF models is studied { a
factor that has often been neglected in spatial hearing research. In the far-eld,
at a distance larger than approximately 1 meter, distance eects can generally be
neglected, but at closer ranges they play an important yet fairly unexplored role
in localization. Related literature can be found from, e.g., (Rabinowitz et al.,
1993; Brungart and Rabinowitz, 1996; Duda and Martens, 1997; Brungart et al.,
1997; Calamia and Hixson, 1997; Brungart, 1998). In the experiments carried out
by the author (Huopaniemi and Riederer, 1998), near-eld and far-eld HRTFs
were measured and compared to analytical HRTF models. In the following, the
experiment and model results are presented for the ITD and ILD.
The HRTF measurements were performed in the large anechoic chamber of the
Laboratory of Acoustics and Audio Signal Processing at the Helsinki University of
Technology (Riederer, 1998a). For the far-eld measurements, Audax AT100M0
4-inch elements in plastic 190 mm spherical enclosures attached to an aluminum
framework (covering seven elevation angles) were placed at a 2.0 m distance. In
the near-eld measurement, a smaller sound source (LPB 80 SC 3-inch element
in a plastic 150 mm spherical enclosure) was at a distance of 0.7 m to the subject.
For this study, measurements were carried out at 0 elevation and 10 azimuth
increments. The spherical shape of the sound sources was considered as the

80

5. Binaural Modeling and Reproduction

most optimal for a single clement loudspeaker casing with minimum diraction.
The size and shape of the element used for near-eld measurements was found
not to aect the measurement result at the distance of 0.7 m; at a closer range
the point-source approximation would not have been valid anymore3 . A total of
eight human subjects and one dummy head (Cortex MK II) were measured in
this study in both near-eld and far-eld setups. The test person was placed in
a measurement chair that was xed to a turntable. The ear canals of the subject
were blocked with moldable silicon putty, to which two Sennheiser KE-41 l-2
miniature microphones were attached. The measurements were fully computer
controlled.
Investigating the ITD as a Function of Distance

The question of whether the ITD changes as a function of distance has been
addressed in several publications (Brungart and Rabinowitz, 1996; Duda and
Martens, 1997). In order to verify the results presented elsewhere, a simple study
was conducted using the 9 measured subjects' HRTFs and comparing those ITDs
to the ones computed using a structural model (presented in Section 3.2.1). The
measurement results were compared to analytical data based on a spherical head
HRTF model (Rabinowitz et al., 1993; Duda and Martens, 1998). In Fig. 5.8 it
can be seen that the ITD varies only slightly when the source moves from far-eld
to near-eld, as has been stated in previous articles (Brungart et al., 1997; Duda
and Martens, 1998).
Investigating the ILD as a Function of Distance

As previously shown and discussed in the literature (Brungart and Rabinowitz,

1996; Duda and Martens, 1997), the ITD cue dierences as a function of distance
(except for very close distances) are not perceptually important. It is, however,
not the case in ILD behavior.
Results for modeling HRTFs from two observation distances in the free-eld
are shown in Figs. 5.9{5.10 (based on (Duda and Martens, 1998)). From these
gures it can be clearly seen that the lower frequencies appear boosted in the
ipsilateral (same) side and damped in the contralateral (opposite) side at a close
distance when compared to far-eld. At higher frequencies the tendency is similar.
In these gures the angle values (on the right side of the plots) are those of the
incidence from the source to the observation point (the ear) on the surface of the
sphere (not to be confused with absolute azimuth angle used elsewhere). In order
to study the ILD measurements as a function of distance, the following procedure
for the HRTFs was carried out. For near-eld and far-eld measurements on
3

Note that the near-eld eects in HRTFs at the distance of 0.7 m are not expected to be
signicant. This has been shown by (Duda and Martens, 1998; Brungart, 1998). This distance
was chosen due to practical reasons caused by the measurement setup.

5.1

81

Properties and Modeling of HRTFs

ITD for distances 1.9 and 0.7 m

x 10

HRTF Farfield
HRTF Nearfield

ITD (s)

5
0

50

100

150

200

250

300

350

Azimuth (deg)
4

x 10

Spherical Farfield
Spherical Nearfield

ITD (s)

5
0
5
0

50

100

150
200
Azimuth (deg)

250

300

350

Figure 5.8: ITD plots in far-eld and near-eld for HRTFs measured on human
subjects (top graph, mean of 9 subject ITDs) and a spherical head model (bottom
graph).
human subjects, the mean of the magnitude response for both cases at 30 degree
azimuth angle intervals was calculated. The ILD was thus computed using the
following formula
ILDazi (! ) = 20 log

PN
i=1

jt [hrtfr;azi(i; !)]j 20 log PNi=1 jt[hrtfl;azi(i; !)]j

N
N

(5.15)
where N is the number of measured subjects (in this case, N = 9), azi is the
azimuth angle, and ! is the frequency variable. In Figs. 5.11{5.14, ILD results
based on measurements are depicted and compared to those based on a Rayleigh
diraction model. In the upper plots, the ILDs for near-eld and far-eld for the
measured HRTFs and the analytical model are plotted. In the lower plots, the
change in ILD as the source moves to the near-eld is plotted for the measured
HRTFs and the analytical model. From the gures an increase in ILD can be
seen at closer distance when the source moves to the side (in both ipsilateral and
contralateral cases). The trend of ILD increase can be seen in the gures. As the
source comes closer, the ILD dierences increase rapidly as shown by Brungart
(1998); Duda and Martens (1998). The analytical model is shown to predict the
ILD increase well (when moving the source from the 2.0 m distance to 0.7 m).

82

5. Binaural Modeling and Reproduction

Spherical head HRTF for a source at 0.3 m distance

10

0
50

70

Magnitude Response (dB)

90
100

-5

110

-10

180
130

-15

140

-20

150
170

-25
-30
10

10
Frequency (Hz)

10

Figure 5.9: HRTFs computed from a Rayleigh diraction model (Duda and
Martens, 1998). Source distance: 0.3m.
Spherical head HRTF for a source at 2.0 m distance
10
0

Magnitude Response (dB)

90
100

110

-5

180

-10

130
140

-15

150
160

-20

170

-25
10

10
Frequency (Hz)

10

Figure 5.10: HRTFs computed from a Rayleigh diraction model (Duda and
Martens, 1998). Source distance: 2.0m.

83

Properties and Modeling of HRTFs

ILD comparison, mean across 9 subjects, azi=0 deg

Magnitude (dB)

30

HRTF Farfield
HRTF Nearfield
Spherical Farfield
Spherical Nearfield

20

10

10
2
10

10

10

Magnitude (dB)

10

HRTF Fartonearfield gain

Spherical Fartonearfield gain

10
2
10

10

10

Frequency (Hz)

Figure 5.11: Near-eld and far-eld interaural level dierences at 0.
ILD comparison, mean across 9 subjects, azi=30 deg
Magnitude (dB)

30

20

10

10
2
10

10

10

10

Magnitude (dB)

5.1

10
2
10

10

10

Frequency (Hz)

Figure 5.12: Near-eld and far-eld interaural level dierences at 30.

84

5. Binaural Modeling and Reproduction

ILD comparison, mean across 9 subjects, azi=60 deg

Magnitude (dB)

30

20

10

10
2
10

10

10

Magnitude (dB)

10

10
2
10

10

10

Frequency (Hz)

Figure 5.13: Near-eld and far-eld interaural level dierences at 60.
ILD comparison, mean across 9 subjects, azi=120 deg
Magnitude (dB)

30

20

10

10
2
10

10

10

Magnitude (dB)

10

10
2
10

10

10

Frequency (Hz)

Figure 5.14: Near-eld and far-eld interaural level dierences at 120.

5.1

Properties and Modeling of HRTFs

85

5.1.5 Functional Modeling of HRTFs

Apart from a signal-modeling approach to HRTFs, interest in functional representations of HRTFs has also risen over the past years in search of ecient
auralization techniques. These methods are not directly related to specic lter
design issues, but can serve as a basis for, e.g., structural smoothing and preprocessing of the data. One of the main goals in this approach is the decomposition
of HRTFs to components that enable computational or memory reduction in
the synthesis stage. Principal components analysis (PCA) is one of the bestknown decomposition methods used for reorganizing data. In the decomposition,
the data is reorganized into basis functions and weight vectors that produce an
ordered model of the eective variance of the sample space. Using this decomposition, perfect resynthesis of the data can be accomplished. Remarkable data
reduction can also be achieved if only the rst basis functions (explaining often
90-95% of the variance) are used.
Principal components analysis (PCA) for HRTF modeling was rst proposed
by Martens (1987). It has been since been studied by, e.g., Kistler and Wightman
(1992); Middlebrooks and Green (1992); Blommer (1996). Kistler and Wightman
(1992) used PCA to approximate minimum-phase diuse-eld HRTFs (that they
called directional transfer functions, DTFs). In this method the magnitude spectra of the HRTFs were approximated using ve principal spectral components of
the response. With this method the order of the resulting FIR lters was successfully reduced to 1=3 of the original impulse response with only a slight decrease
in localization accuracy.
Chen et al. (1995) have proposed a feature extraction method, where a complex valued HRTF was represented as a weighted sum of eigentransfer functions
generated using the Karhunen-Loeve expansion (KLE)4 . The dierence from the
previous PCA model is that a complex HRTF transfer function including magnitude and phase information can be modeled. In the method proposed by Evans et
al. (1998), functional representations for a set of HRTF measurements have been
developed. The HRTFs were represented as a weighted sum of surface spherical harmonics that are a hierarchical set of basic functions orthogonal upon the
surface of a sphere. The reconstruction based on surface spherical harmonics
yielded complex frequency responses that model HRTF data with good accuracy. Computational eciency when compared to the KLE has, however, been
found inferior (Evans et al., 1998). An attractive method suitable for real-time
implementation of virtual acoustic displays (VAD) was presented in (Abel and
Foster, 1997, 1998). In this patent, a method utilizing singular value decomposition (SVD) was used to derive an orthogonal set of functions for a set of HRTFs.
Thus, a linear combination of these functions (basis impulse responses, or complex HRTF basis spectra modeled by digital lters) can be used to synthesize
any HRTF in the set with reasonable computational cost, in a similar manner
4

The PCA and KLE methods are equivalent in this context (Blommer, 1996).

86

5. Binaural Modeling and Reproduction

as proposed by previous studies (e.g., (Martens, 1987; Kistler and Wightman,

1992)). In addition to decomposition methods, structural analysis and modeling
of the HRTF can yield ecient approximations suitable for real-time processing
(Brown and Duda, 1998).
5.2

Auditory Analysis of HRTFs and Application to Filter Design

In this section, binaural processing is approached from the auditory perception

point of view. The properties of human peripheral hearing serve as a starting
point for auditorily motivated binaural signal processing and lter design. It is
proposed by the author that binaural lter design should be carried out using
auditory criteria. Similar conclusions have been made by Kulkarni and Colburn
(1997b).
5.2.1 Properties of the Human Peripheral Hearing
From the psychoacoustical point of view, the linear frequency scale used in linear
transforms such as the Fourier transform is not optimal. In psychoacoustics it
has been shown experimentally that scales such as the Bark scale (or the criticalband rate scale) (Zwicker and Fastl, 1990) or the ERB (Equivalent Rectangular
Bandwidth) rate scale (Moore et al., 1990) match closely the properties of human
hearing. Moreover, presently the ERB scale is believed to be theoretically better
motivated than the Bark scale (Moore et al., 1996). Approximation formulas
for the psychoacoustic scales are the following. The ERB scale bandwidth as a
function of center frequency fc (in kHz) is given by the following equation (Moore
et al., 1990):

fCE = 24:7(4:37fc + 1):
(5.16)
Similarly, the Bark scale can be approximated by the equation (Zwicker and Fastl,
1990):



fCB = 25 + 75 1 + 1:4fc2 0:69 :
(5.17)
In Fig. 5.15, four resolution functions are compared. The plotted functions are
linear, logarithmic, Bark, and ERB rate scales vs. log frequency (Huopaniemi
and Karjalainen, 1997). As can be seen from Fig. 5.15, the log and the ERB
resolution functions are relatively close to each other. The Bark resolution is
similar above 500 Hz. The constant bandwidth resolution function related to the
linear frequency scale is generally not acceptable when characterizing responses
from the auditory point of view. Based on modern psychoacoustic theory (Moore
et al., 1990), a conclusion may be drawn that the auditory resolution is best

5.2

Auditory Analysis of HRTFs and Application to Filter Design

87

Resolution (Q-value)

10

10

10

10

10
Frequency (Hz)

10

Figure 5.15: Frequency resolution (Q-value) curves as functions of frequency:

Solid line = third-octave (constant Q); o-o = ERB resolution; +-+= Bark (critical
band) scale; *-* = constant 100 Hz bandwidth (linear resolution)(Karjalainen et
al., 1999).
represented using the ERB scale, the logarithmic and the Bark scales being useful
approximations, and the linear scale being inferior. The question whether the
monaural psychoacoustic principles apply in the same way to binaural hearing is
an open one. Thus, care should be taken when using (monaural) psychoacoustical
measures in binaural design.
There are dierent methods of incorporating the non-uniform resolution of
hearing into a lter design. The approaches have been divided into three dierent
categories:
 auditory smoothing of the responses prior to lter design
 use of a weighting function in lter design
 warping of the frequency axis prior or during the lter design
In the following, these three approaches will be discussed.
5.2.2 Auditory Smoothing
In many cases, the data (in this case the HRTF frequency-domain or time-domain
responses) may be preprocessed using a smoothing technique prior to the lter

88

5. Binaural Modeling and Reproduction

design (Abel, 1997). The most straightforward way to accomplish a frequencydependent magnitude response smoothing is to use moving averaging of variable
window size (Smith, 1983; Koring and Schmitz, 1993):
s

jHS (f )j = f f
1
0

Z f1
f0

jH (f 0)j2 df 0;

(5.18)

where f1p f0 is the window

p width at frequency f 0. Furthermore, if we dene
f0 = f= K and f1 = f K , it follows that for third-octave width, K = 5=4
and for octave width, K = 2 (Koring and Schmitz, 1993). The window size
as a function of frequency can also be derived using auditory criteria, based on
the Bark or ERB scale. In these smoothing techniques the window widths at
frequencies f 0 are derived using Eqs. (5.16{5.17).
Another technique for HRTF smoothing that has been considered is cepstral
smoothing (Huopaniemi et al., 1995; Kulkarni and Colburn, 1997a). This technique is widely used in speech processing (Oppenheim and Schafer, 1989) and
is also closely related to minimum-phase reconstruction of impulse-responses via
cepstral analysis, described in Section 5.1. Cepstral smoothing can be understood
as linear smoothing of the log magnitude response. Because it does not involve
a non-uniform frequency resolution, it is not strictly an \auditory smoothing"
technique. In Figures 5.16{5.17, the eects of dierent smoothing techniques applied to an HRTF magnitude response are illustrated. The HRTF data in these
gures are based on real head measurements (Riederer, 1998a).
5.2.3 Auditory Weighting
An alternative to auditory smoothing techniques is to use weighting functions in
the lter design that approximate the human auditory resolution. Approximation
formulas presented in Eqs. (5.16) and (5.17) have been used to calculate the
weighting functions shown in Fig. 5.18. The weight of each frequency point is
the inverse of the bandwidth calculated with Eqs. (5.16) and (5.17). As can be
seen from Fig. 5.18, the ERB scale weighting function focuses more at low and
high frequencies (below 500 and above 4000 Hz) than the corresponding Bark
function, but remains very close to the Bark scale at mid-frequencies.
5.2.4 Frequency Warping
It has been proposed earlier in this section that modeling and lter design of
HRTFs should be carried out using psychoacoustical criteria. One attractive
possibility to approximate a non-linear frequency resolution is to use frequency
warping. Approximations of HRTFs using frequency warping have not been extensively studied. Jot et al. (1995) have proposed a method where the HRTFs
are preprocessed using auditory smoothing and the IIR lter design is carried

5.2

Auditory Analysis of HRTFs and Application to Filter Design

89

HRTF: person: 1, azimuth: 0, elevation: 0, right ear

10
0

Relative Magnitude (dB)

10
20
30
40
50
256tap orig
Bark
ERB
1/3 oct
1/10 oct
Cepstral

60
70
80
2

10

10

10

Frequency (Hz)

Figure 5.16: Dierent smoothing techniques applied to HRTF magnitude response approximation.
out using a Yule-Walker algorithm (Friedlander and Porat, 1984) in the warped
frequency domain5. A framework for warped binaural lter design has been established by the author (Huopaniemi and Karjalainen, 1996b,a, 1997; Huopaniemi et
al., 1998, 1999b). The fundamentals of warped lters are studied in the following.
Frequency scale warping is in principle applicable to any design or estimation
technique. The most popular warping method is to use the bilinear conformal
mapping (Strube, 1980; Smith, 1983; Smith and Abel, 1995; Karjalainen et al.,
1997b; Smith and Abel, 1999). The bilinear warping is realized by substituting
unit delays with rst-order allpass sections:
D1 (z ) =

(5.19)
1 z 1 ;
where  is the warping coecient. This implies that the frequency-warped version
of a lter can be implemented by such a simple replacement technique. It is easy
to show that the inverse warping (unwarping) can be achieved with a similar
substitution but using  instead of  (Smith, 1983; Jot et al., 1995). Warping
curves for dierent values of  are illustrated in Fig. 5.19.
z

This procedure was originally suggested in (Smith, 1983) and applied for modeling of the
violin body.

90

5. Binaural Modeling and Reproduction

HRIR: person: 1, azimuth: 0, elevation: 0, right ear

1

Amplitude

6
0

10

20

30

40
50
60
Time (samples)

70

80

90

Figure 5.17: Minimum-phase time-domain HRIRs of smoothed magnitude responses shown in Fig. 5.16.
The usefulness of warping by conformal mapping comes from the fact that,
given a target transfer function H (z), it may be possible to nd a lower order
warped lter Hw (D1(z)) that is a good approximation of H (z) in an auditory
sense. Hw (D1 (z)) is then designed in a warped frequency domain so that using
allpass delays instead of unit delays maps the design to a desired transfer function
in the ordinary frequency domain. For an appropriate value of , the bilinear
warping can t the psychoacoustic Bark scale, based on the critical band concept
(Zwicker and Fastl, 1990), surprisingly accurately (Smith and Abel, 1995). An
approximate formula for the optimum value of  as a function of sampling rate is
given in (Smith and Abel, 1995)6. For a sampling rate of fs = 48 kHz this yields
 = 0:7364, and for fs = 32 kHz  = 0:6830 (Smith and Abel, 1999).
5.3

Digital Filter Design of HRTFs

The task of approximating an ideal HRTF response H (z) by a digital lter H^ (z)
is studied in this section. For any lter design, the goal is to minimize an error
6 Revised

version and formula has been published in (Smith and Abel, 1999).

5.3

91

Digital Filter Design of HRTFs

ERB weighting
Bark weighting
2

Relative Weight

10

10

10

10

10
Frequency (Hz)

10

Figure 5.18: Auditory weighting functions as a function of frequency, fs = 48

kHz.
function that is of the form:

E (ej! ) = H (ej! ) H^ (ej! )

(5.20)
A list representing research carried out in the eld of HRTF approximation by
various authors is illustrated in Table 5.1. The lter order values have been
collected from the corresponding publications and represent FIR lter orders
(number of taps - 1) or IIR lter orders (number of poles or zeros). One can see
from the table that the results from dierent studies vary considerably. There
are many reasons for this. Some of the studies are purely theoretical, meaning
that the results are formulated in the form of a spectral error measure or by
visual inspection. In some of the references, the authors claim that a certain
lter order appeared to be satisfactory in informal listening tests. These cases
are marked in the table with a question mark. There have been very few formal
listening tests in this eld that also give statistically reliable results (Sandvad
and Hammershi, 1994a; Huopaniemi and Karjalainen, 1997; Kulkarni and Colburn, 1997a,b). Another question is the validity of the HRTF data used in the
studies. The HRTFs may have been equalized for free-eld conditions or a certain headphone type, or the data may have have been based on dierent HRTF
measurements (dummy-head, individual/non-individual data). Dierent preprocessing such as minimum-phase reconstruction may also have been applied. All

92

5. Binaural Modeling and Reproduction

0.
8
0.
6

0.
4

0.4

0.2
0

0.
2

0.6

0
= .0

0.
=
2
0.

4
=
-0
.6

=
-0
.8

0.8

Normalized warped frequency

0.2
0.4
0.6
0.8
Normalized original frequency

Figure 5.19: Frequency warping characteristics of the rst-order allpass section

D1 (z ) for dierent values of . Frequencies are normalized to the Nyquist rate
(Karjalainen et al., 1999).
these aspects could cause the large deviation seen in the results of the studies
(Table 5.1).
5.3.1 HRTF Preprocessing
The sound transmission in an HRTF measurement includes characteristics of
many subsystems that are to be compensated in order to achieve the desired
response. The transfer functions of the driving loudspeaker, the microphone,
and the ear canal (if the measurement position were inside an open ear canal)
may then have to be equalized. If, however, a more general database of HRTFs
is desired, other equalization strategies should be considered like equalization
with respect to a given reference direction, or diuse-eld equalization. In general, HRTF preprocessing strategies can be divided into (Huopaniemi and Smith,
1999):
 Temporal preprocessing
 Spectral preprocessing
By temporal preprocessing we normally consider minimum-phase reconstruction and separate modeling and implementation of the ITD (discussed in Section 5.1.1). Spectral preprocessing comprises auditory spectral preprocessing
(smoothing, weighting, and warping as discussed in the previous section) and
equalization techniques.

5.3

93

Digital Filter Design of HRTFs

Research Group

Design Type

Filter

Begault, 1991
Sandvad and Hammershoi, 1994ab
Kulkarni and Colburn, 1995, 1997
Hartung and Raab, 1996
Asano et al., 1990
Sandvad and Hammershoi, 1994ab
Blommer and Wakeeld, 1994
Jot et al., 1995
Ryan and Furlong, 1995
Kulkarni and Colburn, 1995, 1997
Kulkarni and Colburn, 1995, 1997

Binaural / FIR
Binaural / FIR
Binaural / FIR
Binaural / FIR
Binaural / IIR
Binaural / IIR
Binaural / IIR
Binaural / IIR
Binaural / IIR
Binaural / IIR
Binaural / IIR

Order

Study

81-512 Empirical
72
Empirical
64
Empirical
48
Empirical
>40
Empirical
48
Empirical
14
Theoretical
10-20 Empirical?
24
Empirical?
6
Empirical?
25 (all- Empirical?
pole)
Hartung and Raab, 1996
Binaural / IIR 34/10 Empirical
Mackenzie et al., 1996
Binaural / IIR 10
Theoretical
Blommer and Wakeeld, 1997
Binaural / IIR 40
Theoretical
Table 5.1: Binaural HRTF lter design data from the literature.
Equalization Methods

Equalization methods of head-related transfer functions have been summarized

in (Mller, 1992; Blauert, 1997; Larcher et al., 1998). The main reason to equalize
the data is to compensate for the response of the measurement or reproduction
systems. For example, one could aim at designing a general database of HRTFs
for diuse-eld equalized headphones. There are two main methods for HRTF
equalization:
 Free-eld equalization
 Diuse-eld equalization
Free-eld equalization is achieved by dividing the measured HRTF by a reference
measured in the same ear from a certain direction. The reference direction is
typically chosen as 0 azimuth and 0 elevation, that is, from the front of the
listener (Mller, 1992; Jot et al., 1995). For clarity, only the magnitude spectrum equalization is considered here; the phase is assumed to be minimum-phase
reconstructible.
jH(!; ; )j = jH (!;jH (=!;0; ; )=j 0 )j

(5.21)

94

5. Binaural Modeling and Reproduction

HRTF: person: 1, azimuth: 20, elevation: 0, right ear

60
50

Original
Df equalized

Relative Magnitude (dB)

40
30
20
10
0
10
20
30 2
10

10

10

Frequency (Hz)

Figure 5.20: Illustration of unequalized and diuse-eld equalized HRTF magnitude responses.
where  is the azimuth angle,  is the elevation angle, and ! is the angular
frequency. Diuse-eld HRTFs, on the other hand, are estimated by poweraveraging all HRTFs from each ear. The equalized HRTFs are obtained by dividing the original measurement by the diuse-eld HRTF of that ear.
jHdf(!; ; )j = q

jH (!; ; )j
jHi(!; ; )j2

1 PM
i=1
M

(5.22)

where M is the total number of HRTFs in the set. In the diuse-eld normalized
responses a attening of the HRTF spectra is normally achieved. This is due
to the removal of spectral factors that are not incident-angle dependent (such
as the ear canal resonance, if present in the measurements). In Fig. 5.20, the
dierence between a normal (measurement equalized) HRTF and a diuse-eld
equalization is illustrated. Flattening of the spectra and reduced dynamic range is
clearly visible. This suggests that the needed lter order for perceptually correct
design should also be lower, as will be seen in further sections.

5.3

95

Digital Filter Design of HRTFs

5.3.2 Error Norms

Let us look at the problem of nding an optimal lter approximation to a given
HRTF. A common way to quantify errors in digital lter design is to use the Lpnorms (Golub and Loan, 1996; Smith, 1983). An Lp error norm of a frequencydomain representation of an arbitrary lter H (ej! ) and its approximation H^ (ej! )
is dened by the following equation:
kE kp = kH ( ) H^ (ej! )kp ,
ej!

Z  

H ej!


p
( ) H^ (ej! ) d!
2

1

;p  1

(5.23)

Least-Squares Norm

The L2 norm is generally known as the least squares (integrated squared magnitude) norm. It is particularly appealing due to the fact that by using Parseval's
relation this norm can directly be used both in the frequency and the time domains (Smith, 1983; Parks and Burrus, 1987):
E2 = kH (ej! ) H^ (ej! )k22
Z 
2

j!
j!
^
, H (e ) H (e ) d!
2
(5.24)

1


X
2
= h(n) h^ (n) , kh(n) h^ (n)k22
n=0

If a positive weighting function W (ej! ) is further employed, a general weighted

least squares (WLS) complex approximation problem in the frequency domain
can be written as:
Z 
2


d!
j!
j!
j!
^
Ew2 ,
W e H (e ) H (e )
(5.25)
2

This property is desirable since it is possible to incorporate weighting functions
into the lter design and optimization in order to model, e.g., the auditory frequency resolution. Another advantage of the least squares formulation is that
since the error term is quadratic, there is a global minimum. Popular methods
using least-squares error norm minimization include the equation-error method,
Prony's method, and the Yule-Walker methods (Smith, 1983).
Chebyshev Norm

The L1 norm is often referred to as the Chebyshev norm. This norm, as p

approaches innity, aims to minimize the maximum component of the error term
E (hence the term minimax):



j!
j!
j!
j!
^
^
kE k1 = kH (e ) H (e )k1 , <!<
max H (e ) H (e )
(5.26)

96

5. Binaural Modeling and Reproduction

The Chebyshev norm minimization may be an applicable choice for modeling

HRTFs, since the magnitude responses typically exhibit prominent peaks and
valleys that are important for localization. Since the minimax approach seeks to
minimize the maximum error, which is often found at the transitions in the response, good performance in HRTF modeling would be expected. Another property of human hearing, the logarithmic amplitude resolution, can also be incorporated in a Chebyshev norm minimization algorithm (a weighted log-magnitude
spectral error (Smith, 1983)). A drawback of the Chebyshev norm optimization
is, however, that the error surface may not always be convex and the result can be
unstable and/or the search process only locally optimal. Well-known methods for
Chebyshev norm optimization are, e.g., the Remez multiple exchange algorithm
(used in the Parks-McClellan algorithm) and simplex methods (Steiglitz, 1996).
Hankel Norm

Hankel norm based methods are attractive for HRTF modeling since they provide a general and stable solution to complex frequency response modeling. The
Hankel norm lies between L2 and L1 norms and is quantied by the spectral
norm of the Hankel matrix of a given response. The most straightforward technique for obtaining an optimal Hankel norm approximation is to nd the Hankel
singular values of a Hankel matrix (by singular value decomposition) and select
a desired number of the singular values to construct the nal lter (Smith, 1983;
Mackenzie et al., 1997). In this work, two Hankel norm lter design methods
have been studied: balanced model truncation (BMT) (Mackenzie et al., 1997)
and the Caratheodory-Fejer (CF) method (Gutknecht et al., 1983).
5.3.3 Finite Impulse-Response Methods
The most straightforward ways to approximate HRTF measurements are to use
the window method for FIR lter design or a direct frequency sampling design
technique (Parks and Burrus, 1987). If H (ej! ) is the desired frequency response,
the direct frequency sampling method is carried out by sampling N points in the
frequency domain and computing the inverse discrete Fourier transform (normally
using the FFT):
h(n) =

1 NX1 H (ej!k )ej!k n;

k=0

(5.27)

where !k = 2N k; and k = 0; : : : ; N 1. A time-domain equivalent to uniform

frequency-sampling is windowing (truncating) the sampled impulse response h(n)
with a rectangular window function w(n). Generally, the window method can be
presented as a multiplication of the desired impulse response with some window

5.3

Digital Filter Design of HRTFs

function w(n):

97

h^ (n) = h(n)w(n):

(5.28)
The use of the rectangular window is known to minimize a truncated time-domain
L2 norm (Oppenheim and Schafer, 1989). By Parseval's relation, then, the frequency response is also optimal in the unweighted least squares sense over the
spectral samples !k . The eect of dierent window functions in HRTF lter design has been discussed by Sandvad and Hammershi (1994a). They concluded
that although rectangular windowing provokes the Gibbs' phenomenon, seen as
ripple around amplitude response discontinuities, it is still favorable when compared to, e.g., the Hamming window. Another and perhaps more intuitive conclusion is based on the fact that HRTFs do not contain spectral discontinuities but
rather broad uctuations in the magnitude response, which can be modeled eectively using least squares tting. A severe limitation of the windowing method is,
however, the lack of a frequency-domain weighting function that could model the
non-uniform frequency resolution of the ear. For an extended frequency-sampling
method yielding an LS solution, it is possible, however, to introduce non-uniform
frequency sampling and weighting (Parks and Burrus, 1987).
Kulkarni and Colburn (1995a, 1997a) have proposed the use of a weighted least
squares technique based on log-magnitude error minimization for nite-impulse
response HRTF lter design. They claim that an FIR lter of order 64 is capable
of retaining most spatial information.
In a method proposed by Hartung and Raab (1996), binaural lters were optimized using auditory criteria, approximating the sensitivity of the human ear
with a logarithmic magnitude weighting function. Non-uniform sampling of the
frequency grid was applied in order to achieve auditory resolution. Optimization
was then carried out to yield the following results: an FIR lter of order 48 \revealed no signicant dierences in localization performance... Minor divergence
was noted with the 32nd-order FIR lter" (Hartung and Raab, 1996).
Issues in FIR lter design using auditory criteria have been discussed by
Wu and Putnam (Wu and Putnam, 1997). They derived a perceptual spectral
distance measure using a simplied auditory model. The technique was applied
to HRTF-like magnitude responses and as a result FIR approximations of order
20 were successfully calculated.
5.3.4 Innite Impulse-Response Methods
The earliest HRTF lter design experiments using pole-zero models were carried
out by Kendall and Rodgers (1982); Kendall and Martens (1984). A comparison of FIR and IIR lter design methods has been presented in (Sandvad and
Hammershi, 1994a,b). The non-minimum-phase FIR lters based on individual HRTF measurements were designed using rectangular windowing. The IIR

98

5. Binaural Modeling and Reproduction

lters were generated using a modied Yule-Walker algorithm that performs leastsquares magnitude response error minimization. The low-order t was enhanced
a posteriori by applying a weighting function and discarding selected pole-zero
pairs at high frequencies. Listening tests showed that an FIR of order 72 equivalent to a 1.5 ms impulse response was capable of retaining all of the desired
localization information, whereas an IIR lter of order 48 was needed for the
same localization accuracy.
In the research carried out by Blommer and Wakeeld (1994), the error criteria
in the auto-regressive moving average (ARMA) lter design were based on logmagnitude spectrum dierences rather than magnitude or magnitude-squared
spectrum dierences. Furthermore, a new approximation for the log-magnitude
error minimization was dened. The theoretical study concluded that it was
possible to design low-order HRTF approximations (the given example used 14
poles and zeros) using the proposed method. In more recent studies (Blommer,
1996; Blommer and Wakeeld, 1997), the results have been generalized, and they
concluded that pole-zero models of order 40 were needed for accurate modeling
of HRTFs. They also stated that a least squares (LS) algorithm was inferior in
comparison to using a logarithmic error measure.
Asano et al. (1990) have investigated sound localization in the median plane.
They derived IIR models of dierent orders (equal number of poles and zeros)
from individual HRTF data using the least-squares error criterion (equation-error
method). When compared to a reference, a 40th-order pole-zero approximation
yielded good results in the localization tests with the exception of increased frontback confusions in frontal incident angles.
Other IIR approximation models for HRTFs have been presented by Ryan
and Furlong (1995); Jenison (1995); Hartung and Raab (1996); Kulkarni and
Colburn (1995b, 1997b). In the following, a novel IIR modeling technique based
on balanced model truncation (Mackenzie et al., 1997) is discussed.
Balanced Model Truncation

In this section, a new low-order lter design technique for HRTFs is presented.
An attractive technique for HRTF modeling has been proposed in (Mackenzie
et al., 1997). By using this balanced model truncation (BMT) it is possible to
approximate HRTF magnitude and phase response in a Hankel norm sense with
low order IIR lters (down to order 10). A complex HRTF system transfer
function is written as a state-space dierence function (see, e.g., (Proakis and
Manolakis, 1992)), which is then represented in balanced matrix form (Moore,
1981). A truncated state-space realization Fm (z) corresponds to the original
system transfer function F (z) with a similarity to the original system which is
quantied by the Hankel norm:
kF (z) Fm (z)kH  2tr(2 );
(5.29)

5.3

99

Digital Filter Design of HRTFs

where



0
1
 = 0 2 ;
(5.30)
where 2 = diag(1 ; : : : ; k ) are the Hankel singular values (HSVs) of the rejected system after truncation and 1 = diag(k+1; : : : ; n) are the HSVs of the
truncated system.
A practical and straightforward implementation of BMT has been outlined in
(Beliczynski et al., 1992) and will be presented in the following 7 . The state-space
dierence equations for an FIR lter F (z) = c0 + c1z 1 +


+ cnz n = c0 + F1(z)
can be written as
x(k + 1) = Ax(k) + Bu(k)
(5.31)
y (k) = Cx(k) + Du(k);
where
2
0 0 : : : 0 03
61 0 : : : 0 07
7;
A=6
4
5
:::
0 0 ::: 1 0
2 3
1
(5.32)
607
7
B=6
6 .. 7 ;
4.5
0
C = [c1 c2 c3 : : : cn ]; D = c0 :
The Hankel matrix is formed from the FIR lter coecients F1(z) in the following
way8 :
2

c1 c2 : : : cn
6c c : : : 0 7
2
3
7
H=6
6 ..
.
.
.
.
4.
. : : : . 75
cn 0 : : : 0

(5.33)

Since H is a symmetric matrix, the HSVs can be found and ordered using the
singular value decomposition method (SVD) (Golub and Loan, 1996):
H = V V T
(5.34)
7

Matlab

algorithm

for

BMT

http://www.acoustics.hut.fi/software/hrtf/.
8

implementation

is

available

at

There is an error in (Beliczynski et al., 1992) for the Hankel matrix formulation. The
equation given in Eq. (5.33) is correct.

100

5. Binaural Modeling and Reproduction

where  is a diagonal matrix containing the HSVs and V V T = I is a unit matrix.

In practice, it is desirable to plot the HSVs and determine the desired order of
approximation. In (Beliczynski et al., 1992), the following formulas are presented
for deriving a balanced truncated system of order k from matrix V :
Ak = V (2 : n; 1 : k)T V (1 : n 1; 1 : k)
Bk = V (1; 1 : k)T
(5.35)
Ck = CV (1 : n; 1 : k)
D = c0 ;
where \:" is interpreted as in the Matlab \colon operators" (1 : n , 1; 2; ::; n).
Now it is possible to write the truncated system in state-space representation,
which can then be conveniently transposed to a traditional system transfer function form. Although the above formulation is only an approximation to the
Hankel norm, it has been argued in the literature (Kale et al., 1994) that in cases
where wideband tting is applied, a BMT outperforms in terms of frequency
response error an optimal Hankel-norm design (Chen et al., 1992).
In the experiments carried out in (Mackenzie et al., 1997), minimum-phase,
diuse-eld equalized, auditory smoothed HRTFs (based on Kemar measurements by Gardner and Martin (1994, 1995)) were modeled by 10th-order IIR lters created using BMT. The signal-to-error power ratios (SER) were compared
to IIR models designed using Prony's method (Parks and Burrus, 1987) and the
Yule-Walker method (Friedlander and Porat, 1984). The average SER was found
to be approximately 10 dB better in BMT models (see Fig. 2 in (Mackenzie et
al., 1997)).
5.3.5 Warped Filters
In this section, issues in warped lter design are discussed9 . The underlying
theory of frequency warping was presented in Section 5.2.4.
Warped FIR Structures

A warped FIR lter (WFIR) may be interpreted as an FIR structure in which the
unit delays have been replaced by rst-order allpass lters. Its transfer function
is given by
w (z ) =
9

M
X
i=0

i [D1 (z )]i ;

(5.36)

A Matlab toolbox for frequency warping and implementation of warped lters is available
at http://www.acoustics.hut.fi/software/warp/.

5.3

101

Digital Filter Design of HRTFs

(a)

in
x0

out

D1(z)

x2

z -1 0

x1
D1(z)

(b)

in

z -1

z -1

etc.

out

etc.

Figure 5.21: WFIR structures (Karjalainen et al., 1999). a) Principle of warped

FIR lter. b) Implementation used in present study (Karjalainen et al., 1997a).
in +

y1

+
2

z -1
etc.

z -1

2 x
2
y3

z -1

1 x
1
y2

a)

x0

out

in

b)

g=1/0
1 z -1

out

2 z -1

3 z -1
etc.

Figure 5.22: WIIR structures (Karjalainen et al., 1999). a) Unrealizable direct

form of WIIR lter. b) Realizable modied implementation used in present study.
where D1(z) is as dened in Eq. (5.19). A more detailed lter structure for
implementation is depicted in Fig. 5.21 (Strube, 1980). It can be seen that a
warped FIR structure is actually recursive, i.e., an IIR lter with M poles at
z = 1, where M is the order of the lter. A straightforward method to get the
tap coecients i for a WFIR lter is to warp the original HRTF impulse response
using Eq. (5.19), and to truncate by rectangular windowing the portion that has
the amplitude above a threshold of interest.

102

5. Binaural Modeling and Reproduction

Warped IIR Structures

Having the warped HRTF response at hand, a traditional lter design method
such as Prony's method (Parks and Burrus, 1987) may be applied to yield a
warped pole-zero model of the form:
PM
i [D1 (z )]i
i
=0
Hw (z ) =
:
(5.37)
1 + PRi=1 i [D1 (z)]i
This cannot be implemented directly due to delay-free recursive propagation
through D1 (z) units (Strube, 1980). By proper mapping of i coecients to new
i coecients the warped IIR lter can be implemented as shown in Fig. 5.22
(Imai, 1983; Karjalainen et al., 1997a).
5.4

Binaural System Implementation

Practical binaural lter design and implementation requires knowledge of what

the system will feature. When comparing dierent lter design and implementation strategies, one should pay attention particularly to the following questions:
 Is the system dynamic, i.e., is HRTF interpolation and commutation needed?
 Are minimum-phase and pure-delay HRTF approximations being used?
 Is specialized hardware (signal processors) for implementation being used?
 Are there memory constraints for storing HRTF data?
 Is the HRTF processing part of a room acoustics modeling scheme?
In this work, the goal has been to present and compare methods for HRTF
lter design; thus, the solutions to the above questions are only outlined. In
many dynamic virtual acoustic environments, minimum-phase FIR approximations have been chosen for HRTF implementation due to straightforward interpolation, relatively good spectral performance, and simplicity of implementation.
Furthermore, this approach may easily be integrated into real-time geometric
room acoustics modeling schemes, such as the image-source method (Savioja et
al., 1999). One of the drawbacks is, however, the large memory requirement that
this approach produces. A PCA-based method may be attractive when conducting HRTF spatialization, since as few as 5 principal component basis functions
have been found to model human HRTFs (or DTFs, directional transfer functions) with good accuracy (Kistler and Wightman, 1992).
The following benchmarks have been calculated for a Texas Instruments TMS320C3x oating point signal processor, but are practically similar in other processors as well. FIR implementation is ecient (N+3 instructions for N taps), and

5.4

Binaural System Implementation

103

dynamic coecient interpolation is possible. Designs are usually straightforward

(e.g., frequency sampling), but give limited performance, especially at low orders.
IIR implementations are generally slower (2N+3 instructions for order N direct form II implementations), especially if dynamic synthesis is required. Interpolation and commutation methods, such as cross-fading, and dierent transient
elimination techniques, increase computation. Pole-zero models are suited for
arbitrarily shaped magnitude-response designs; thus, low-order designs are possible.
A second-order section decomposition of IIR lter coecients may be desirable
in many implementations:
QM=2
1
2
Bi (z )
k=1 b0k + b1k z + b2k z
Hsos(z ) = Ki
=
(5.38)
2
1+a z 2
Ai (z ) QN=
1
+
a
z
1
k
2
k
k=1
In Figs. 5.21{5.22, direct (warped domain) implementations for WFIR and WIIR
lters were illustrated. A possible alternative for low-order implementations of
warped lters (high orders may be unstable due to computational precision problems) is to unwarp the warped structures to traditional direct-form pole-zero lter
structures. One such solution is presented in the following. If frequency warping
is used prior to lter design, the warped second-order section implementation can
be unwarped to regular second-order sections by the following substitutions:
QM=2
^
^ 1 ^ 2
Bwi (z )
k=1 b0wk + b1wk z + b2wk z ;
=
(5.39)
Hdwsos =
QN=2
Awi (z )
^1wk z 1 + a^2wk z 2
k=1 1 + a
where
^b0wk = b0wk + b1wk + 2b2wk
=a^0wk
^b1wk = 2b0wk + (1 + 2 )b1wk + 2b2wk
=a^0wk
^b2wk = 2b0wk + b1wk + b2wk
=a^0wk
(5.40)
a^0wk = 1 + a1wk + 2 a2wk


a^1wk = 2 + (1 + 2 )a1wk + 2a2wk =a^0wk


a^2wk = 2 + a1wk + a2wk =a^0wk;
where  is the warping coecient.
The eciency of warped vs. non-warped lters depends on the processor that
is used. For Motorola DSP56000 series signal processors, a WFIR takes three
instructions per tap instead of one for an FIR. For WIIR lters, four instructions
are needed per pole instead of two for an IIR lter. In custom-designed chips,
the warped structures may be optimized so that the overhead due to complexity
can be minimized. The warped structures may also be expanded (unwarped) into
direct form lters, which will lead to the same computational demands as with
normal IIR lters.

104

5. Binaural Modeling and Reproduction

5.4.1 Interpolation and Commutation of HRTF Filters

The task of interpolation and commutation of HRTF lters is an important issue
when designing dynamic real-time virtual audio environments. In (Jot et al.,
1995), the term interpolation has been dened as synthesizing an intermediate
transfer function from a database of predened lters. By commutation (changing) of the coecients of a lter we simply mean dynamic coecient updating.
The topic has been studied in publications by, e.g., Wenzel and Foster (1993);
Runkle et al. (1995); Jot et al. (1995). In the following, methods for FIR and
IIR lter coecient interpolation and commutation (for real-time applications)
are overviewed.
FIR Interpolation

For non-recursive FIR lters, direct coecient interpolation of HRIRs is possible.

Minimum-phase approximations used in conjunction with a delay line implementation for ITD have been found to be perceptually indistinguishable from original
(mixed-phase) HRIRs (Kistler and Wightman, 1992). However, HRIRs are not
completely minimum-phase, and interpolation on mixed-phase lter coecients
may result in a comb-ltering eect in the frequency domain due to changing
delay properties. Interpolation of FIR lter coecients in dynamic auralization
can be expressed both in two- and three-dimensional cases.
For 2-D interpolation the coecients hc(n) for a desired azimuth angle are
obtained from measured HRIRs using the following formula:
hC (n) = (1 c )hA (n) + c hB (n);
(5.41)
where hA and hB are hC 's two neighboring data points. Similarly, in a threedimensional case, the coecients are obtained using 4-point bilinear interpolation
from the four nearest available data points (Begault, 1994). Since the HRTFs
are minimum-phase FIRs this interpolation can be carried out. The interpolation
scheme for point E located at azimuth angle  and elevation  is:
hE (n) = (1 c )(1 c )hA (n) + c (1 c )hB (n)+
(5.42)
c c hC (n) + (1 c )c hD (n);
where hA; hB ; hC ; and hD are hE 's four neighboring data points as illustrated
in Fig. 5.23, c is the azimuth interpolation coecient ( mod grid )=grid, and n
goes from 1 to the number of taps of the FIR lter. The elevation interpolation
coecient is obtained similarly c = ( mod grid)=grid.
The task of interpolating non-minimum-phase FIR approximations of HRIRs
has been found to produce severe comb-ltering eects if the phase delays of
the interpolating lters vary considerably. This phenomenon may occur if nonminimum-phase HRTFs measured at sparse azimuth intervals (> 10 15) are
interpolated to reproduce intermediate angles.

5.4

105

Binaural System Implementation

11
00

11
00
00
11

11
00
00
11

1
0

hC
11
00
hD 0
00
11
1
0
1
hE

00
11
11
00
11
00
00
11
hB
0
1
hA

Figure 5.23: HRTF-lter coecients corresponding to point hE at azimuth  and

elevation  are obtained by bilinear interpolation from measured data points hA,
hB , hC , and hD (Begault, 1994; Savioja et al., 1999).
IIR Interpolation

Dynamic realization of recursive IIR structure interpolation is more complicated.

In (Jot et al., 1995), the possibilities for realizing dynamic 3-D sound synthesis
using IIR lters were discussed. Linear interpolation of two stable second-order
section- or lattice-realization of IIR lters is guaranteed to be stable. In order
to achieve regularity in interpolation, a special pairing and ordering algorithm
applied globally to the lter database was presented in (Jot et al., 1995). In that
study, four possible choices for IIR lter representation were discussed:





direct-form coecients of the second-order sections in cascade

magnitudes and log-frequencies of the poles and zeros
lattice lter coecients ki
log area ratios log[(1 ki) = (1 + ki)]

The process of dynamically updating (commuting) the coecients of recursive

lters may cause transients that are audible as disturbing clicks. This problem
may be dealt with using dierent techniques. The method of cross-fading uses
two lters in parallel, and the output is calculated as a weighted sum (linear interpolation) of the outputs from the lters. The transition time is dened as the

106

5. Binaural Modeling and Reproduction

time it takes to completely change the output from one lter to the other. Typically, in HRTF processing, the lters are chosen to be the azimuth and elevation
states before and after the transition. This method doubles the computational
cost of IIR implementation. In (Valimaki, 1995; Valimaki and Laakso, 1998), a
transient elimination method has been presented, which is based on state variable
updates of the lter coecients. This method reduces the computational cost of
commutation by 50 %. Various methods for transient elimination in time-varying
recursive lters have been discussed in (Valimaki, 1995).
5.5

Objective and Subjective Evaluation of

HRTF Filter Design Methods

This section presents methods and procedures for objective and subjective evaluation of binaural lter design. It is clear that the performance of binaural systems
has to be evaluated using human subjects in formal listening experiments. In addition, it is desirable to have objective measures of performance quality that can
be used for comparing dierent lter designs or implementation techniques. The
objective measures that have been considered in this study are based on modeling
of auditory perception.
5.5.1 Objective Methods
Spectral Distance Measure

There is a need to have a simple numerical measure of lter design quality that is
meaningful also from the perceptual point of view. A similar approach has been
considered in, e.g., (Wu and Putnam, 1997). In (Huopaniemi and Karjalainen,
1997), a spectral (magnitude) distance measure was experimentally derived in
the following way.
The equalized impulse response is rst FFT transformed to magnitude-square
of the frequency response, sampled (by interpolation) uniformly on a logarithmic
frequency scale, smoothed with about 0.2 octave resolution (this resolution value
was specied somewhat arbitrarily to be not too far from the ERB resolution),
and converted to dB scale. In the next step the dierence of the spectrum to
be analyzed and a reference spectrum is computed for the passband region of
approximation. The reference spectrum may be simply the average level of the
spectrum to be analyzed or some other reference. In our case it was the corresponding reference response in our listening experiments. A root-mean-square
value of the dierence spectrum is then computed, and this is used as an objective
spectral distance measure to characterize the perceptual dierence between the
magnitude responses or a deviation from a reference response. Notice that the

5.5

Objective and Subjective Evaluation of

HRTF Filter Design Methods

IACC
IACC

107

ITD
spectrum

GTFB
HRTF L

IACC
pink
noise

GTFB

LL

Left LL
Spectrum

LL

HRTF R

LL
half wave
rectification

low pass
filtering

LL
LL

Right LL
Spectrum

LL

Figure 5.24: Detailed view of the binaural auditory model used in the objective
evaluation of HRTF lter designs (Pulkki et al., 1999).
values of the spectral distance measure used in our study are not calibrated to
be compared directly with any perceptual dierence measures.
Binaural Auditory Model

A novel idea of using a binaural auditory model for virtual source quality estimation has been proposed by Pulkki et al. (1998, 1999)10 This method has been
adopted successfully for binaural lter quality estimation by the author in (Pulkki
et al., 1998; Huopaniemi et al., 1998; Pulkki et al., 1999; Huopaniemi and Smith,
1999; Huopaniemi et al., 1999b). The basis of the auditory model applied in
this study lies in the coincidence and cross-correlation principles introduced by
Jeress (1948). The early model has been further extended by several authors
(e.g. Lindemann (1986) and Gaik (1993)). A schematic of the binaural auditory
model used in this study is depicted in Fig. 5.24 (slightly modied from (Pulkki
et al., 1998)). The model consists of the following steps.
 A pink noise sample convolved with the HRTFs under study is used as

the excitation. Pink noise yields a spectrally balanced excitation to the

10

Similar techniques have been presented by Macpherson (1991).

108

5. Binaural Modeling and Reproduction

auditory system lacking major temporal attacks, thus the in uence of the
precedence eect is minimized.
 The HRTF-ltered pink noise samples are passed through a gammatone
lterbank (GTFB) (Slaney, 1993) of N bandpass ERB (equivalent rectangular bandwidth) channels (N depends on the sampling frequency fs).
 Half-wave rectication and lowpass ltering (cuto frequency at 1 kHz) are
used to simulate the hair cells and the auditory nerve behavior in each
bandpass channel.
For the ITD model, the interaural cross-correlation function (IACC) is calculated
for each bandpass channel pair (see Fig. 5.24). The ITD as a function of ERB
channel is estimated by calculating the time delay corresponding to the position of
the maximum in each bandpass IACC curve. Loudness estimates (L in sones) for
the
p left2 and right ERB bandpass channels are calculated using the equation L =
< H > (an approximation of Zwicker's formula, where he used an exponent
of 0.23 (Zwicker and Fastl, 1990)), where < H 2 > is the average power of the
ERB channel output. The loudness levels LL (in phons) for each ERB channel
are computed using the formula LL = 40+10 log2 L, resulting in a loudness level
spectrum for left and right ear signals.
4

Error Criteria for Binaural Auditory Model

In order to be able to compare the binaural auditory model outputs for dierent
lter designs, a suitable error measure had to be considered. The root-meansquare (RMS) error was found attractive for comparing the binaural auditory
model outputs of dierent lter designs (and lter order) with a reference response. In the calculation of the distance measures, the basic phenomena found
in human sound localization (Blauert, 1997) were observed: 1) the spectral and
interaural level dierence (ILD) cues dominate localization at frequencies above
approximately 1.5 kHz, but may also contribute to localization at lower frequencies, and 2) the interaural time dierence (ITD) is the dominant localization cue
at frequencies below approximately 1.5 kHz. Based on these assumptions, two
quality measures were derived:
 Perceived loudness level spectrum error
 Perceived loudness level spectrum + ITD modeling error
The modeling errors are calculated as an RMS dierence between the outputs
from the auditory models (loudness level spectra or ITDs) of the reference and
the lter approximation over a passband fl fh. In the results presented below
the following limits were chosen: fl = 1:5 kHz and fh = 16 kHz. The loudness
level spectrum error was calculated by summing the left and right ear loudness

5.6

Experiments in Binaural Filter Design and Evaluation

109

errors11 . The loudness level spectrum + ITD modeling error was calculated by
scaling and summing the low frequency (f < 1:5 kHz) ITD modeling error with
the high frequency loudness level spectrum modeling error. Since the role of lowfrequency spectral cues in sound source localization is ambiguous, a choice was
made to exclude the loudness level error at f < 1:5 kHz from the RMS error
measures.
5.5.2 Subjective Methods
The more traditional and straightforward approach (albeit more time-consuming)
for evaluating the quality of binaural systems is to use subjective listening experiments. During the course of this work, several listening experiments were
conducted on the perception of HRTF lter quality degradation and localization
of virtual sound sources in headphone listening.
5.6

Experiments in Binaural Filter Design and

Evaluation

In the following, results from four dierent studies carried out by the author on
binaural lter design and evaluation are presented. First, design examples for
the papers (Huopaniemi and Karjalainen, 1996b,a) are overviewed. Second, the
design examples and listening tests presented in (Huopaniemi and Karjalainen,
1997) are discussed. The third study (Huopaniemi et al., 1998, 1999b) presents
methods and results of individualized HRTF lter design and listening tests.
Finally, the fourth study (Huopaniemi and Smith, 1999) discusses the role of
spectral and temporal preprocessing and equalization methods in binaural lter
design. The goal in these experiments has been to investigate subjective and
objective evaluation of HRTF lter design methods and to determine the validity
of auditory-based lter design methods proposed by the author. Furthermore, the
task has been to nd benchmarks for binaural synthesis, that is, thresholds for
HRTF lter approximation where no perceptual dierences can be found when
compared to original (raw) HRTFs.
5.6.1 Binaural Filter Design Experiment
In (Huopaniemi and Karjalainen, 1996b,a), the application of warped lter design methods to HRTF modeling was tested using measurements carried out on
a dummy head (Bruel&Kjaer 4100). HRTFs were measured for both ears at 5
azimuth intervals in an anechoic chamber using the DRA Laboratories' MLSSA
11

Summation of left and right ear loudnesses to obtain a binaural loudness estimate was
used in (Moore et al., 1997).

110

5. Binaural Modeling and Reproduction

system. Additional measurements were carried out to compensate for the measurement equipment and to obtain headphone correction lters (four headphone
types were used). The internal sampling rate of the MLSSA system was 75.47
kHz and the results were bandpass ltered from 20 to 25000 Hz. In addition,
HRTF measurements made on a Kemar dummy head were modeled (Gardner
and Martin, 1994, 1995). The BK4100 HRTF measurements were preprocessed
according to the following principles. These stages included 1) minimum-phase
reconstruction and 2) smoothing of the magnitude response data. For comparison, diuse-eld equalized Kemar dummy head HRTFs were used.
Frequency warping was performed after preprocessing. Two values of  were
used (0.65 and 0.7233), the rst of which is slightly lower than for approximate
Bark-scale warping, and the latter being close to the optimal Bark-scale match
(for fs = 44:1 kHz). In the warped frequency domain, dierent FIR and IIR lter
design methods were compared. A time-domain IIR design method, Prony's s
method (Parks and Burrus, 1987) was used, because it outperformed other tested
methods (i.e. the Yulewalk method and discrete least-squares tting) especially
in low-order approximations.
In Fig. 5.25, the modeling results for Kemar are illustrated. The example
HRTF was measured at 0 elevation, 30 azimuth, and diuse-eld equalization
was used. It can be seen that a warped Prony design (denoted WIIR2) easily
outperforms a linear Prony design (denoted IIR) of equivalent order for frequencies up to 9 kHz, which is expected. In this case, the order of the IIR lters was
20, compared to the FIR designs of order 40. The frequency-sampling method
was used both in the WFIR and FIR designs. The performance of a WFIR order
13 was compared to a non-warped FIR or order 40. Accuracy is slightly reduced
in WFIR implementation. The performance of WIIRs when compared to the
IIR are superior in both examples at lower frequencies. The tradeo at higher
frequencies is tolerable according to psychoacoustical theory. Furthermore, the
BMT lter design method presented in section 5.3.4 was compared to the previous methods. It can be seen that a very low-order BMT model (order 10) retains
most of the spectral features of the HRTF, although not at the same accuracy as
a WIIR of same order.
In Fig. 5.26, the BK4100 HRTFs were used. The lter orders for modeling were higher than in the previous experiment. The results are similar; the
eciency of the WIIR structure at lower frequencies with a tradeo at high frequencies can be seen clearly. If the WIIR lters are implemented by inverse
warping (unwarping) the coecients into traditional direct form representations,
the WIIR appears to be the best solution. There are constraints in inverse warping caused by numerical accuracy, but in this study WIIR approximations of
orders 10-32 have been successfully unwarped.

5.6

111

Experiments in Binaural Filter Design and Evaluation

Modeling of DF-equalized minimum-phase KEMAR HRTFs: elev=0, azi=30

90

Relative Magnitude (dB)

80
70
60
50
Minimum-phase Original
FIR, window: boxcar, order: 40
WFIR, lambda=0.65, window: boxcar, order: 13
IIR, design: Prony, order: 20
WIIR, lambda= 0.65, design: Prony, order: 10
WIIR2, lambda= 0.7233, design: Prony, order: 20
BMT IIR, order: 10

40
30
20
2

10

10
Frequency (Hz)

10

Figure 5.25: Warped lter design compared with traditional FIR and IIR designs.
Kemar dummy head HRTFs were used (Huopaniemi and Karjalainen, 1996b,a).
Modeling of minimum-phase B&K4100 HRTFs: elev=0, azi=30
20

Relative Magnitude (dB)

10

-10

-20
Minimum-phase Original
FIR, window: boxcar, order: 80
WFIR, lambda=0.65, window: boxcar, order: 27
IIR, design: Prony, order: 40
WIIR, lambda= 0.65, design: Prony, order: 20
WIIR2, lambda= 0.7233, design: Prony, order: 40

-30

-40
2

10

10
Frequency (Hz)

10

Figure 5.26: Warped lter design compared with traditional FIR and IIR designs.
BK4100 dummy head HRTFs were used (Huopaniemi and Karjalainen, 1996b,a).

112

5. Binaural Modeling and Reproduction

5.6.2 Evaluation of Non-Individualized HRTF

Performance
In the second experiment (Huopaniemi and Karjalainen, 1997), the HRTFs used
in the lter design examples and listening tests were measured from a Cortex MK2
dummy head in an anechoic chamber. The Cortex dummy head was equipped
with Bruel&Kjaer 4190 microphones (blocked ear canal version). The transducer
used in the measurements was a four-inch Audax AT100M0 loudspeaker mounted
in a plastic ball. A random-phase at amplitude spectrum pseudorandom noise
signal was used as the excitation sequence. Data were played and recorded using
an Apple Macintosh host computer and the QuickSig signal processing environment (Karjalainen, 1990). A National Instruments NI-2300B DSP card with
Texas Instruments TMS320C30 signal processor and high-quality 16-bit AD/DA
converters were used for both HRTF measurements and listening experiments.
The HRTF data were post-processed for headphone listening experiments
in the following way. A compensation measurement was made to account for
the measurement system by placing a microphone at the dummy head position
with the head absent. Headphone transfer functions for the Sennheiser HD580
headphone were measured on the dummy head. A 300-tap FIR inverse lter for
each ear was designed using least-squares approximation. The HRTF data was
then convolved by the compensation response and the headphone correction lter
for both ears.
The minimum-phase reconstruction was carried out using windowing in the
cepstral domain (as implemented in the Matlab Signal Proc. Toolbox rceps.m
function (Mathworks, 1994)). The cross-correlation method was used to nd the
ITD for each incident angle. The ITD was inserted as a delay line. Three different minimum-phase HRTF approximations were used: windowed FIR design
(rectangular window), time-domain IIR design (Prony's method (Parks and Burrus, 1987)) and a warped IIR design (warped Prony's method, warping coecient
 = 0:65. In Table 5.2, the processed lter lengths for dierent lter types are
collected. The example HRTFs were measured at 0 and 135 azimuth (0 elevation) positions. The magnitude responses of the HRTF lter approximation using
IIR and WIIR design for 135 azimuth can be seen in Figs. 5.27{5.28. Again, it
can be seen that a warped Prony design has an improved low-frequency t (up
to approximately 9 kHz) when compared to a linear Prony design of equivalent
order. The better t at low frequencies when comparing WIIR approximation to
windowed FIR can also be observed. The value of  = 0:65 was used, which is
slightly lower than for approximate Bark-scale warping. Figure 5.29 plots the
spectral distance measures as functions of the number of lter coecients for the
three HRTF lter types used in our study: FIR, IIR, and WIIR. The spectral
distance measure has been calculated using the method described in Section 5.5.1
for HRTFs from both ears at the two azimuth angles (0 and 135) used in the
listening test. The computational load from the implementation point of view

5.6

Experiments in Binaural Filter Design and Evaluation

FIR (rect. wind.)

IIR (Prony's method)

Warped IIR (

113
 = 0:65)

256 (reference)
128
128
128
64
64
96
48
48
88
44
44
80
40
40
72
36
36
64
32
32
60
30
30
56
28
28
52
26
26
48
24
24
44
22
22
40
20
20
36
18
18
32
16
16
28
14
14
24
12
12
20
10
10
16
8
8
Table 5.2: HRTF lter types and orders used in the second listening experiment.
of these structures is comparable, provided that the WIIR lters are unwarped
to traditional direct form IIRs. The reference for distance computation was the
highest-order response, in order to make the results compatible to the setup used
in our listening experiments.
Listening Test Procedure

In order to verify the theoretical lter design results, headphone listening experiments were carried out. The goal in the study was to detect the just noticeable
dierence (JND) thresholds of audibility by varying the lter order and comparing it to a reference HRTF (similarly as in (Sandvad and Hammershi, 1994a,b)).
Listening experiments were carried out for the three HRTF lter design methods
described in the previous section: FIR, IIR, and WIIR. A total of 8 test subjects
participated in the listening experiment, 6 male and 2 female with ages ranging
between 21 and 35. The hearing of all test subjects was tested using standard
audiometry. None of the subjects had reportable hearing loss that could eect
the test results. It should be pointed out here that the experiment was done using
non-individualized HRTFs (measured on a dummy head) that were equalized for
a specic headphone type (Sennheiser HD580).

114

5. Binaural Modeling and Reproduction

IIR approximation, azimuth=135, elevation 0, left ear

Original 256-tap FIR

Relative Magnitude (dB)

-10

48

-20
36

-30
-40

24

-50
12

-60
-70
-80 2
10

10
Frequency (Hz)

10

Figure 5.27: IIR lter design of Cortex dummy head HRTFs (135 azimuth). IIR
lter orders: 12, 24, 36, 48 (Huopaniemi and Karjalainen, 1997).
In the listening experiment the TAFC (Two Alternatives Forced Choice)
bracketing method was used. The method is described in great detail in (ISO,
1989) and widely used in, e.g., audiometric tests. In each trial, two test sequences were presented with a 0.5 s interval between the samples. The rst test
signal was always the reference signal, and the second signal varied according to
adaptation. Each test type was played three times and only the last two were
accounted for in the data analysis. The test subjects were given written and oral
instructions. They were also familiarized with a test sequence that demonstrated
both distinguishable and indistinguishable test signal pairs.
A total of four dierent stimuli were rst processed for a pilot study; pink
noise, male speech, and female speech, and a music sample. All samples were
digitally copied and processed from the Music for Archimedes CD12. In the nal
experiment, however, only the pink noise sample was used. This was due to the
fact that remarkable dierences in dierent lter designs could clearly be heard
only using wide-band test signals. A pink noise sample with a length of one
second (50 ms onset and oset ramps) was used in the nal experiment. The
level of the stimuli was adjusted so that the peak A-weighted SPL did not exceed
70 dB at any point. This has been done in order to avoid level adaptation and
the acoustical re ex (Stapedius re ex).
12

Music for Archimedes, CD B&O 101 (1992).

5.6

115

Experiments in Binaural Filter Design and Evaluation

WIIR approximation, azimuth=135, elevation 0, left ear, lambda=0.65

Original 256-tap FIR

Relative Magnitude (dB)

-10

48

-20
36

-30
-40

24

-50
12

-60
-70
-80 2
10

10
Frequency (Hz)

10

Figure 5.28: WIIR lter design of Cortex dummy head HRTFs (135 azimuth).
WIIR lter orders: 12, 24, 36, 48 (Huopaniemi and Karjalainen, 1997).
The test subject was seated in a semi-anechoic chamber (anechoic chamber
with a hard cardboard oor). The test stimuli were presented over headphones.
A computer keyboard was placed in front of the test subject. Each test subject
was individually familiarized and instructed to respond in the following way:
 Press 1 if the signals are the same
 Press 2 if the signals are dierent
 Press Space if you want to repeat the signal pair
As a total, three dierent lter approximations for two apparent source positions were used. Each alternative was played three times. The results of the
listening tests were gathered automatically by a program written for the QuickSig environment (Karjalainen, 1990). The resulting data were transferred into
Matlab, where analysis was performed.
Results

In Fig. 5.30, the results of the listening test are presented. This gure illustrates
the distribution of just noticeable dierence (JND) thresholds calculated across
two tested azimuth angles, 0 and 135 with three lter types, FIR, IIR, and

116

5. Binaural Modeling and Reproduction

20
18

IIR

Spectral Distance Measure

16
14

FIR

12
10
8
6

WIIR

4
2
0

20

30

40

50
60
70
80
90
Number of Filter Coefficients

100

110

120

Figure 5.29: Characterization of lter design quality using a spectral distance

measure as a function of the number of lter coecients for the three lter types
of the study: FIR, IIR, and WIIR (Huopaniemi and Karjalainen, 1997).
WIIR. The median value as well as the lower and upper quartile (25% and 75%)
values are shown.
The results show that the distribution of results in the listening panel was
relatively small, although not as well dened as a pilot study indicated. This
may be a consequence of using an inhomogeneous listener panel. Some of subjects
were experienced analytic listeners while some did not have prior experience in a
listening panel. A longer training prior to nal experiments could have made the
test results more systematic (Bech, 1993).
From Fig. 5.30 it can be seen that the WIIR performance from the lter
order point of view is superior when compared to FIR and IIR designs. From
the computational point of view, however, the FIR and WIIR implementations
appear to be approximately equal in performance. The warped IIR designs,
however, easily outperforms a conventional IIR design. A useful criterion to
select lter order values could be the upper quartile (75%) or even higher level of
subject reactions. Using the 75% quartile results, one concludes in the following
statements.
For non-individualized (dummy-head) HRTFs equalized to a specic headphone, the lter orders at the sampling rate of 48 kHz where 75% of the popula-

5.6

117

Experiments in Binaural Filter Design and Evaluation

Listening test results: azimuth angles 0, 135

90
FIR

IIR

WIIR

80
70

Filter Order

60
50
40
30
20
10
1

2
HRTF Approximation Type

Figure 5.30: Results of the listening test. The boxplot depicts the median
(straight line) and the 25%/75% percentiles (Huopaniemi and Karjalainen, 1997).
tion stated no dierence when compared to the reference were approximately:
 Order 40 for a frequency-sampling FIR design
 Order 24 for a time-domain IIR design (Prony's method)
 Order 20 for a warped IIR design (Prony's method,  = 0:65)

In comparison to the results presented in the literature, some comments can

be made. The empirical study by Sandvad and Hammershi (1994a,b) resulted
in orders 72 for a FIR and 48 for an IIR lter. The dierence compared to the
results may be caused by the fact that Sandvad and Hammershoi used individual
HRTFs and headphone calibration, and both speech and pink noise samples.
However, the estimated detection probabilities (maximum likelihood estimation
was used as a statistical model) for the given results were approximately 0.08,
higher than in our conclusions. Moreover, the lter orders used in that study
were relatively sparse (orders 24, 36, 72 and 128 for FIR, and orders 10, 20, 30
and 48 for IIR using pink noise).

118

5. Binaural Modeling and Reproduction

Conclusion and Discussion

In the spectral distance measure of Fig. 5.29 it can be seen that from the auditory point of view the WIIR lters have the best performance. As can be seen in
further analysis, The correlation with the listening test results is good. The JND
values for FIR and WIIR lters according to the listening tests were approximately 40 (lter coecients), and for the IIR structure approximately 50. It is,
however, a valid question, why the FIR lters performed so well in the listening
tests without having any weighting or frequency warping applied. This may have
been caused by the use of non-individualized (dummy-head) HRTFs.
5.6.3 Evaluation of Individualized HRTF Performance
This section presents the results of individualized HRTF lter design in objective
and subjective experiments (Huopaniemi et al., 1998, 1999b).
First, the HRTF data and the methods of lter design used in the objective
and subjective analysis of this work are described. The task was to further
investigate three dierent lter design approaches (FIR, IIR, WIIR), and the
goal was to nd methods and criteria for subjective and objective evaluation.
This research concentrated on nding answers to two problems often found in
HRTF lter design:
 What is the needed lter length for perceptually relevant HRTF synthesis?
 Is it possible to introduce an \auditory resolution" in the lter design,
whereby the spectral and interaural phase details are modeled more accurately at low frequencies and considerably smoothed at high frequencies?
A set of 10 human test subjects were chosen for the experiment. A blocked
ear canal HRTF measurement technique (Mller et al., 1995) was used to obtain
the needed transfer functions for the experiments (the used measurement setup
is discussed in greater detail in (Riederer, 1998a). An Audax AT100M0 4-inch
transducer element in a plastic 190 mm enclosure was used as a sound source
for the measurements. The test subjects were seated in an anechoic chamber on
a rotating measurement chair. Sennheiser KE211-4 miniature microphones were
used. HRTF and headphone responses (for the used headphone type: Sennheiser
HD580) were measured for each test person to enable full individual HRTF simulation. Pseudorandom noise was used as the measurement sequence. For the
objective and subjective tests, four azimuth angles were chosen: 0o, 40o, 90o,
and 210o. These incident angles represent the median plane, frontal plane, and
horizontal plane responses.
The minimum-phase reconstruction was carried out using windowing in the
cepstral domain. The minimum + excess phase approximation method (Jot et
al., 1995) was used to nd the ITD for each incident angle. The ITD was inserted in HRTF synthesis as a delay line. Three dierent minimum-phase HRTF

5.6

Experiments in Binaural Filter Design and Evaluation

119

approximations were used: Windowed FIR design (rectangular window), timedomain IIR design (Prony's method (Parks and Burrus, 1987), as implemented
in the Matlab Signal Processing Toolbox (Mathworks, 1994)) and a warped IIR
design (warped Prony's method, warping coecient  = 0:65). The reference
lter order for each incident angle was chosen to be 256 (257 FIR taps). The
tested lter orders were:
 FIR: 96, 64, 48, 32, 16
 IIR: 48, 32, 24, 16, 8
 WIIR: 48, 32, 24, 16, 8
As an example, in Fig. 5.31 magnitude responses of original and lter approximations from one test subject's HRTFs at 40o azimuth (0o elevation) are depicted.
It can be seen that a WIIR design (using Prony's method) has an improved lowfrequency t (up to approximately 9 kHz) when compared to a uniform frequency
resolution IIR design of equivalent order. The better t at low frequencies when
comparing WIIR approximation to windowed FIR can also be observed both in
the amplitude response plots and the ITD plot. The warping value of  = 0:65
was used, which is slightly lower than for approximate Bark-scale warping. This
value was chosen by visual inspection of the magnitude responses to enhance lowfrequency t but still retain the overall high-frequency magnitude envelope. The
results of using a binaural auditory model (discussed in section 5.5.1) in HRTF
lter design quality estimation are depicted in Figs. 5.32{5.34. The loudness
level spectrum estimates of HRTFs for a test subject at 40o azimuth are shown in
Figs. 5.32{5.33 and the perceived ITD of the corresponding lter approximations
is plotted in Fig. 5.34 (note that the loudness level spectrum estimates shown
in Fig. 5.33 correspond to the magnitude responses shown in Fig. 5.31). The
dashed line in each subplot is the reference response, and the solid line is the current approximation result. The number of lter coecients for the row of plots
is shown to the right of the gures. Furthermore, an RMS error estimate of the
corresponding design is shown in Figs. 5.32{5.33 in each of the subplots. From
the plots it can seen that the spectral detail clearly visible in the original HRTF
(see Fig. 5.31) is smoothed as a consequence of auditory modeling. The RMS
error estimate shows that over the passband of loudness level error calculation
(1.5-16 kHz) the warped IIR design method provides the best t to the desired
response. This is also true in the case of the ITD estimate, shown in Fig. 5.34.
The results of estimating HRTF lter design quality using a binaural auditory
model suggest that high-frequency smoothing of the HRTFs is motivated. The
results depicted in Figs. 5.35{5.36 conrm that the RMS loudness level spectrum
error is lowest for the auditory WIIR lter design, which essentially provides
a better low-frequency t with a tradeo in high-frequency accuracy. In the

120

5. Binaural Modeling and Reproduction

Modeling of minimumphase HRTFs: azi=40deg, person: 3

20
15

Relative Magnitude / dB

10
5
0
5
10
15
20
25

Minimumphase Original
FIR, order: 48
IIR, design: Prony, order: 24
WIIR, lambda= 0.65, design: Prony, order: 24

30
10

10

Frequency / Hz

Figure 5.31: HRTF magnitude response and dierent lter approximations. Right
ear, azimuth angle 40o, person 3 (Huopaniemi et al., 1998, 1999b).
gures, both left and right ear HRTF approximations were modeled for 9 test
subjects at 4 azimuth angles. It can be seen that an auditory scale lter (WIIR)
outperforms both FIR and IIR lter design methods. The results are even more
dramatic when the ITD modeling error is added to the error measure. In Fig.
5.36, these results for 9 test subjects at 4 azimuth angles are depicted. In both
plots, the lter design errors start to increase as the number of coecients is
below 48. The WIIR design error is tolerable to order 16, whereas both the FIR
and IIR design errors start to increase earlier.
Subjective Analysis

In order to verify the theoretical lter design results headphone listening experiments were carried out. The goal was to study the performance of individualized
HRTF lter approximations using dierent design techniques and dierent lter
orders. Listening experiments were carried out for the three HRTF lter design
methods as described in the previous section: FIR, IIR, and WIIR. A total of
10 male test subjects participated in the listening experiment, with ages ranging
between 25 and 51. The hearing of all test subjects was tested using standard
audiometry. None of the subjects had reportable hearing loss that could eect

5.6

121

Experiments in Binaural Filter Design and Evaluation

FIR, azi=40, id=3

90
80
70

LL(L) / phons

WIIR, azi=40, id=3

97

3.1908

90
80
70

3.0824

0.96047

65
5.1938

90
80
70

5.8484

1.9814

49
5.4452

90
80
70
90
80
70
0.2

IIR, azi=40, id=3

7.1516

2.433

33
7.1978

8.6572

4.4735

17
9.0677

9.7706

10 21

0.2

1
3
10 21
Frequency / kHz

7.4997

0.2

10 21

Figure 5.32: Binaural auditory model evaluation of lter design quality. Left
ear, azimuth angle 40, person 3. Solid line: lter approximation, dashed line:
reference (Huopaniemi et al., 1998, 1999b). The RMS error is shown on each
graph, and the corresponding number of lter coecients is given on the right
side of the gure.
the test results. The HRTF approximations were individually equalized for a
specic headphone type (Sennheiser HD580).
An A/B paired comparison hidden reference paradigm was employed for the
listening tests with two independent grading scales. The subjects were asked
to grade localization and timbre impairment against the hidden reference on a
continuous 1.0 to 5.0 scale (as proposed in ITU-R BS 1116-1 (ITU-R, 1997)).
The hidden reference in each case was the 257-tap lter. In each trial, two test
sequences were presented with 0.5s between sample (i.e. A/B//A/B). A full
permutation set was employed and two dierent random orders of presentation
were used to minimise bias. To obtain data regarding listener reliability the
reference (the 257-tap FIR) case was also tested against itself. Each test type
was repeated two times. Listeners were given written and oral instructions.
A pink noise sample with a length of one second (50 ms onset and oset ramps)
was used as sound stimulus in the nal experiment. The level of the stimuli was
adjusted so that the peak A-weighted SPL did not exceed 70 dB at any point.
This has been done in order to avoid level adaptation and the acoustical re ex

122

5. Binaural Modeling and Reproduction

FIR, azi=40, id=3

90
80
70

LL(R) / phons

WIIR, azi=40, id=3

97

3.1241

90
80
70

2.4967

0.7613

65
5.0635

90
80
70

5.293

1.7185

49
5.1856

90
80
70
90
80
70
0.2

IIR, azi=40, id=3

6.2493

2.1406

33
6.3121

7.4215

5.0108

17
8.1712

8.2394

10 21

0.2

1
3
10 21
Frequency / kHz

7.3156

0.2

10 21

Figure 5.33: Binaural auditory model evaluation of lter design quality. Right
ear, azimuth angle 40, person 3. Solid line: lter approximation, dashed line:
reference (Huopaniemi et al., 1998, 1999b).
(Stapedius re ex). No gain adjusting of the test sequences calculated for one
person was carried out, since the only variability in level was (possibly) introduced
by the used HRTF lters.
The test procedure in the experiment was as follows. The listener was seated
in a semi-anechoic chamber (anechoic chamber with hard cardboard oor). The
test stimuli were presented over headphones. A computer keyboard was placed
in front of the listener. Each listener was individually familiarized and instructed
to grade the localization and timbre scales for each test signal pair. An example
plot of the listening test software user interface is shown in Fig. 5.37.
A total of ve dierent lter approximations for each of the three lter types
at four apparent azimuth source positions were used. Each alternative was played
three times. The results of the listening tests were gathered automatically by a
program written for the QuickSig environment (Karjalainen, 1990). The result
data were transferred into the SPSS software (Statistical Package for Social Sciences), where analysis was performed.

5.6

123

Experiments in Binaural Filter Design and Evaluation

FIR, azi=40, id=3

IIR, azi=40, id=3

WIIR, azi=40, id=3

0
0.5

97

1
0

ITD / ms

0.5

65

1
0
0.5

49

1
0
0.5

33

1
0
0.5
1
0.2

17

0.5

0.2

0.5
1
Frequency / kHz

0.2

0.5

Figure 5.34: Binaural auditory model evaluation of lter design quality. ITD, azimuth angle 40, person 3. Solid line: lter approximation, dashed line: reference
(Huopaniemi et al., 1998, 1999b).
Results and Discussion

The data were initially checked to ensure equal variance across listeners. At this
point it was found that one listener had very dierent variance than the others.
Upon closer inspection this listener was found to be grading very similarly for all
systems, had non-normal distribution of data and also had very low error variance
and poor F-statistics, based upon an analysis of variance (ANOVA). This was
an indication that this listener could not discriminate between systems and was
thus eliminated from subsequent analysis.
The data were then tested for conformance with the analysis of covariance
model (ANCOVA) assumption. The data were found to be normally distributed,
though slightly skewed, which is typical of subjective data. However, the ANCOVA model is fairly robust to slight skewness of data. Residuals were found
to be normally distributed. It should be noted that the model did not meet the
requirements for homogeneity of variance and the Levene statistics were found
to be signicant. This is not considered problematic as in all other respects the
(ANCOVA) assumptions have been met and the raw and modeled data were
found to be strongly correlated.
An ANCOVA was employed for the full analysis of the data considering all

124

5. Binaural Modeling and Reproduction

Subjects: 9, azims: 4
8

FIR
IIR
WIIR

Composite modeling error

2
10

20

30

40
50
60
70
Number of filter coefficients

80

90

100

Figure 5.35: Binaural auditory model results for 9 subjects and 4 azimuth angles
using three dierent lter design methods. The composite (L + R) loudness level
spectrum error estimate was used (Huopaniemi et al., 1998, 1999b).
factors: lter order (FILTSIZ), lter type (TYPE), listener (PERSON), reproduction angle (ANGLE). A covariate (ORDER) was included to represent the order
(two levels) in which the test was performed. A full factorial analysis was made
for all factors and covariates, employing a type III sum of squares. This analysis
was repeated for both dependent variables: Localization and Timbre. A thorough
discussion of the analysis of covariance results can be found in (Huopaniemi et
al., 1999b). The dominant factor for both variables was that of FILTSIZ. When
considering the 257-tap FIR case, compared against the reference (i.e. itself) it
can be seen that the mean grading is not 5.0, as it should be (Localization: 4.8,
Timbre: 4.6). This is a common phenomenon in listening tests, as often untrained
listeners are not eager to employ the extremes of the grading scale (Stone and
Sidel, 1993, pp. 221-222). However, these values are very close to the reference
and may be within the Just Noticeable Dierence (JND) for this task. Upon
inspection of means as a function or PERSON, it can be seen that listeners are
grading consistently with a common trend (see Figs. 5.38{5.39).
It is clear from Figs. 5.40{5.41 that there is only marginal performance dierence for both variables between FIR and IIR lters for all lter sizes. Clearly, the

5.6

125

Experiments in Binaural Filter Design and Evaluation

Subjects: 9, azims: 4
16

FIR
IIR
WIIR

15

ITD+L+R modeling error

14
13
12
11
10
9
8
7
10

20

30

40
50
60
70
Number of filter coefficients

80

90

100

Figure 5.36: Binaural auditory model results for 9 subjects and 4 azimuth angles
using three dierent lter design methods. The combined ITD and composite (L
+ R) loudness level spectrum error estimate was used (Huopaniemi et al., 1998,
1999b).
17-coecient IIR lter, graded below 3.0 (slightly annoying), should be avoided.
The WIIR design appears to reach an asymptotic level above 4.5 for 49 coecients and above. The FIR and IIR designs only reach this quality level with
the 97-coecient lter. For these designs at least a 65-coecient lter should be
used to exceed the 4.0 (perceptible but not annoying) level. The WIIR lter type
also aords quite appreciable qualities with only 33 coecient lters, providing
grades exceeding 4.0 for both variables.
Considering the second most signicant factor, TYPE, in all cases the WIIR
is found superior, with the FIR design in second place. Based upon the subjective data presented in Figs. 5.40{5.41, obtaining the same level of localization
and timbre quality with individualised HRTFs, requires an FIR lter of approximately double the length of an equivalent quality WIIR lter. The IIR lter
implementation must be even longer to reach this level of quality.
When considering the degradation as a function of ANGLE, it can be seen
from Figs. 5.42{5.43 that timbre degradation is more strongly aected than localization. The highest quality for both scales occurs at 90. This implies that it is
possible to achieve the same quality level for 90 with inferior lters than at other

126

5. Binaural Modeling and Reproduction

Figure 5.37: Listening test questionnaire window for HRTF lter design quality
(localization and timbre grading) (Huopaniemi et al., 1998, 1999b).
angles. Blauert has discussed the human localization insensitivity (localization
blur) in the horizontal plane at the sides, which may provide an explanation for
this phenomenon (Blauert, 1997). It can also be considered that listeners are
more sensitive to timbre, and to a lesser extent localization degradation in the
frontal direction.
In this experiment (Huopaniemi et al., 1998, 1999b), the HRTF lter design
problem was addressed from the objective and subjective evaluation point of view.
Filter design methods taking into account the non-uniform frequency resolution
of the human ear were studied and summarized. A new technique for deriving an
objective HRTF lter design error was incorporated, based on a binaural auditory model. Subjective listening tests were performed to compare the theoretical
model results with empirical localization performance.
The results suggest the following conclusions:
 The high-frequency spectral content present in HRTFs can be smoothed
using auditory criteria without degrading localization performance.
 A binaural auditory model can be used to give a quantitative prediction of
perceptual HRTF lter design performance.

5.6

Experiments in Binaural Filter Design and Evaluation

127

 A warped IIR lter of order 16 (33 coecients) appears to be sucient for

retaining most of the perceptual features of HRTFs in individual subjective

analysis. This conclusion is supported by the objective analysis results.
In conclusion, it can be stated that lter design methods for 3-D sound can gain
considerable eciency when an auditory frequency resolution is used. The nonuniform frequency resolution can be approximated using pre-smoothing, weighting functions, or frequency warping. The binaural auditory model outputs and
listening test results gave similar results in terms of detectable (perceptually audible) dierences in original and approximated HRTFs. The required lter length
for high-quality 3-D sound synthesis is, however, also dependent on the incident
angle of the incoming sound.

128

5. Binaural Modeling and Reproduction

5.0

4.5

95% CI Predicted Value for LOCALIZATION

4.0

3.5

3.0

TYPE
2.5
FIR
2.0
IIR
1.5

WIIR

1.0
1

PERSON

Figure 5.38: Listening test results for three lter types (FIR, IIR, WIIR).
Predicted values as a function of test persons. Tested variable: localization
(Huopaniemi et al., 1998, 1999b).
5.0

4.5

4.0

95% CI Predicted Value for TIMBRE

3.5

3.0

TYPE
2.5
FIR
2.0
IIR
1.5

WIIR

1.0
1

PERSON

Figure 5.39: Listening test results for three lter types (FIR, IIR, WIIR). Predicted values as a function of test persons. Tested variable: timbre (Huopaniemi
et al., 1998, 1999b).

5.6

129

Experiments in Binaural Filter Design and Evaluation

5.0

4.5

95% CI Predicted Value for LOCALIZATION

4.0

3.5

3.0

TYPE
2.5
FIR
2.0
IIR
1.5

WIIR

1.0
17

33

49

65

97

257

FILTSIZ

Figure 5.40: Listening test results for three lter types (FIR, IIR, WIIR). Predicted values as a function of lter type and order. Tested variable: localization
(Huopaniemi et al., 1998, 1999b).
5.0

4.5

4.0

95% CI Predicted Value for TIMBRE

3.5

3.0

TYPE
2.5
FIR
2.0
IIR
1.5

WIIR

1.0
17

33

49

65

97

257

FILTSIZ

Figure 5.41: Listening test results for three lter types (FIR, IIR, WIIR). Predicted values as a function of lter type and order. Tested variable: timbre
(Huopaniemi et al., 1998, 1999b).

130

5. Binaural Modeling and Reproduction

5.0

4.5

95% CI Predicted Value for LOCALIZATION

4.0

3.5

3.0

TYPE
2.5
FIR
2.0
IIR
1.5

WIIR

1.0
0

40

90

210

ANGLE

Figure 5.42: Listening test results for three lter types (FIR, IIR, WIIR). Predicted values as a function of presentation angle. Tested variable: localization
(Huopaniemi et al., 1998, 1999b).
5.0

4.5

4.0

95% CI Predicted Value for TIMBRE

3.5

3.0

TYPE
2.5
FIR
2.0
IIR
1.5

WIIR

1.0
0

40

90

210

ANGLE

Figure 5.43: Listening test results for three lter types (FIR, IIR, WIIR).
Predicted values as a function of presentation angle. Tested variable: timbre
(Huopaniemi et al., 1998, 1999b).

5.6

Experiments in Binaural Filter Design and Evaluation

131

5.6.4 Eects of HRTF Preprocessing and Equalization

This section presents the results of preprocessing and equalization in HRTF lter
design in objective and subjective experiments (Huopaniemi and Smith, 1999).
First the HRTF data, the methods of lter design, and the objective analysis
tool shall be described. The task was to investigate the behavior of dierent
lter designs with varying preprocessing. The goal was to determine and verify
the quality of reproduction using objective evaluation.
Description of HRTF Data

A set of 5 human test subject HRTFs were chosen as data for the experiment (subjects AFW, SJX, SOS, SOU, SOW from the Wightman&Kistler HRTF
database). HRTFs from the right ear of each subject at 30 azimuth intervals
( 150 : 30 : 180) and 0 elevation were used.
The minimum-phase reconstruction was carried out using windowing in the
cepstral domain (as discussed in Section 5.1.1). The minimum + interaural excess
phase approximation method was used to nd the ITD for each incident angle.
The ITD was inserted in HRTF synthesis as a non-fractionally addressed delay
line. Two sets of HRTFs were derived: the original measurements (loudspeaker
and microphone responses deconvolved) and the diuse-eld equalized versions
(using Eq. 5.22).
In the pilot test six dierent minimum-phase HRTF approximations, six different smoothing techniques, and three dierent lter design orders were used:
 Filter Order: 24, 12, 8
 Filter Type: Baseline comparison (FIR windowing method13), Yulewalk
(Mathworks, 1994), Prony (Mathworks, 1994), Invfreqz (Mathworks, 1994),
Steiglitz-McBride (Mathworks, 1994), BMT (Mackenzie et al., 1997), CF
(Gutknecht et al., 1983)
 Smoothing Type: No smoothing, Bark, ERB, 1/3-octave, 1/10-octave, Cepstral (see 5.2.2 for details on the smoothing algorithms)
 Equalization: No equalization, diuse-eld equalization
In the nal experiment and statistical analysis of the results, all previous methods
except the CF and cepstral smoothing were present.
In order to be able to compare the binaural auditory model outputs (Section
5.5.1) for dierent lter designs, a suitable error measure had to be dened. A
\monaural loudness-level spectrum error" has been dened as the root-meansquare (RMS) dierence between the loudness level (computed by the binaural
13

The lter order for the baseline FIR lters is double to that of the IIR lters in order for
the comparison to be relevant. Thus the corresponding FIR orders are 48, 24, and 16.

132

5. Binaural Modeling and Reproduction

Model: Yulewalk, order: 8, person: 1, azimuth: 20, elevation: 0, right ear

60

40

21.8778

Relative Magnitude (dB)

21.5842

20

21.6527
21.3595

21.2569

20

19.696

256tap orig
Bark
ERB
1/3 oct
1/10 oct
Unsmoothed
Windowed FIR

40

60

80
10

10
Frequency (Hz)

10

Figure 5.44: Yulewalk IIR modeling, order 8, auditory model error shown in
graphs (Huopaniemi and Smith, 1999).
auditory model) of a reference HRTF (256-tap FIR) and that of an approximating
lter. The designed lters varied according to spectral smoothing used, equalization, lter type, and lter order. Examples of auditory model output for two
lter design methods (Yulewalk, BMT) and dierent smoothing techniques are
illustrated in Figs. 5.44{5.45.
Results and Discussion

The lter designs and binaural auditory model simulations were carried out in
Matlab (Mathworks, 1994). A total of 9360 lters were designed from the HRTF
database of 5 human subjects. Statistical analysis and plotting was conducted
using SPSS. The results are graphically summarized in Figs. 5.46{5.51. The
results show clearly that diuse-eld equalized lters generally provide a lower
perceptual error, and the error variance across HRTF designs is smaller. This is
due to attening of the spectra as discussed in previous sections. Furthermore,
the baseline lter design (Windowed FIR) seems to be outperformed by other
design methods, especially the Hankel norm based BMT. (The CF method performed similarly to the BMT, but not quite as good, so it is not shown in the
gures.) It is notable that by choosing proper preprocessing techniques (namely,
minimum-phase reconstruction, diuse-eld equalization and auditory smoothing), the choice of lter design methods becomes a non-crucial task, although
clear dierences can still be found between methods.

5.6

133

Experiments in Binaural Filter Design and Evaluation

Model: BMT, order: 8, person: 1, azimuth: 20, elevation: 0, right ear

60

40

Relative Magnitude (dB)

14.2365
16.3287

20

16.3926
16.6194

17.9536

20

19.696

256tap orig
Bark
ERB
1/3 oct
1/10 oct
Unsmoothed
Windowed FIR

40

60

80
10

10
Frequency (Hz)

10

Figure 5.45: BMT IIR modeling, order 8, auditory model error shown in graphs
(Huopaniemi and Smith, 1999).
In this study (Huopaniemi and Smith, 1999), methods for temporal and spectral preprocessing of binaural digital lters were presented. The perceptual quality of dierent preprocessing and smoothing schemes was compared using a simple
binaural auditory model. The results suggest the following conclusions:
 controlled smoothing applied to HRTFs results in better (psychoacoustically motivated) magnitude tting
 diuse-eld equalization allows for lower order lter designs
 Hankel norm optimal algorithms (BMT) perform slightly better at lower
lter orders when compared to other tested methods, but in general proper
choices of preprocessing diminish the dierence between lter designs and
error norms.
It should be noted that the objective validation of the methods was only
performed on the magnitude response. Since all designed lters were minimumphase, the phase characteristics are likely to have little variation across designs.
Nevertheless, more detailed auditory analysis of combined temporal and spectral
preprocessing modeling aspects will be performed in the future. Furthermore, the
perceptual validity of diuse-eld equalization should be veried in subjective
listening experiments. The validity of the objective binaural auditory model
in comparing the dierent designs can be questioned, but in a previous study

134

5. Binaural Modeling and Reproduction

(Huopaniemi et al., 1998, 1999b) the model was found to correspond well to
subjective listening experiment results.
5.7

Discussion and Conclusions

In this chapter, the principles of spatial hearing and binaural synthesis were rst
presented. The author has contributed to binaural lter design by introducing
new methods based on modeling the human auditory frequency resolution. New
methods proposed by the author based on frequency warping and balanced model
truncation were discussed. Furthermore, techniques and experiments on objective
and subjective evaluation of binaural lter design were presented. Novel methods
estimating the quality of HRTF lters have been created by the author, enabling
objective and subjective evaluation of binaural systems. Discussion and results
of four experiments carried out by the author were presented.
The main conclusion of this chapter is that the computational eciency of
binaural processing can be improved by taking into account the auditory resolution in the lter design stage. This nding is supported by theoretical analysis
and objective metrics as well as by subjective listening experiments.
In the next chapter, aspects in crosstalk canceled binaural lter design and
implementation are discussed.

5.7

135

Discussion and Conclusions

U n e q u a liz e d H R T F s , F ilte r O r d e r : 8
3 0

2 0

W in d o w e d F IR
N o s m o o th in g

P e rc e p tu a l E rro r

1 0

B a rk
E R B
1 /3 -o c ta v e

1 /1 0 -o c ta v e
Y u le w a lk

In v fre q z
P ro n y

B M T
S te ig litz - M c B r id e

Figure 5.46: Design of unequalized HRTFs, order 8 (Huopaniemi and Smith,

1999).
U n e q u a liz e d H R T F s , F ilte r O r d e r : 1 2
3 0

2 0

W in d o w e d F IR
N o s m o o th in g

P e rc e p tu a l E rro r

1 0

B a rk
E R B
1 /3 -o c ta v e

1 /1 0 -o c ta v e
Y u le w a lk

In v fre q z
P ro n y

B M T
S te ig litz - M c B r id e

Figure 5.47: Design of unequalized HRTFs, order 12 (Huopaniemi and Smith,

1999).

136

5. Binaural Modeling and Reproduction

U n e q u a liz e d H R T F s , F ilte r O r d e r : 2 4
3 0

2 0

W in d o w e d F IR
N o s m o o th in g

P e rc p e tu a l E rro r

1 0

B a rk
E R B
1 /3 -o c ta v e
1 /1 0 -o c ta v e

Y u le w a lk

In v fre q z
P ro n y

B M T
S te ig litz - M c B r id e

Figure 5.48: Design of unequalized HRTFs, order 24 (Huopaniemi and Smith,

1999).
D iffu s e - fie ld H R T F s , F ilte r O r d e r : 8
3 0

2 5

2 0

D F W in d F IR

1 5

D F N o s m o

P e rc e p tu a l E rro r

1 0

D F B a rk
D F E R B

D F 1 /3 -o c t
0

D F 1 /1 0 -o c t
Y u le w a lk

In v fre q z
P ro n y

B M T
S te ig litz - M c B r id e

Figure 5.49: Design of DF-equalized HRTFs, order 8 (Huopaniemi and Smith,

1999).

5.7

137

Discussion and Conclusions

D iffu s e - fie ld H R T F s , F ilte r O r d e r : 1 2

3 0

2 5

2 0

D F W in d F IR

1 5

D F N o s m o

P e rc e p tu a l E rro r

1 0

D F B a rk
D F E R B

D F 1 /3 -o c t
0

D F 1 /1 0 -o c t
Y u le w a lk

In v fre q z
P ro n y

B M T
S te ig litz - M c B r id e

Figure 5.50: Design of DF-equalized HRTFs, order 12 (Huopaniemi and Smith,

1999).
D iffu s e - fie ld H R T F s , F ilte r O r d e r : 2 4
3 0

2 5

2 0

D F W in d F IR

1 5

D F N o s m o

P e rc e p tu a l E rro r

1 0

D F B a rk
D F E R B

D F 1 /3 -o c t
0

D F 1 /1 0 -o c t
Y u le w a lk

In v fre q z
P ro n y

B M T
S te ig litz - M c B r id e

Figure 5.51: Design of DF-equalized HRTFs, order 24 (Huopaniemi and Smith,

1999).

138

5. Binaural Modeling and Reproduction

Chapter 6
Crosstalk Canceled Binaural
Reproduction
The idea of presenting crosstalk-compensated binaural information over a pair of
loudspeakers was introduced almost 40 years ago (Bauer, 1961), and rst formulated into practice by Schroeder and Atal (Schroeder and Atal, 1963; Atal and
Schroeder, 1966). They described the use of a crosstalk cancellation lter for
converting binaural recordings made in concert halls for loudspeaker listening.
Their impressions of listening to loudspeaker reproduced dummy-head recordings were \nothing less than amazing". However, they observed the limitations
of the listening area, the \sweet spot", which has remained an unsolved problem
in loudspeaker binaural reproduction ever since.
The psychophysical and acoustical basis for manipulating stereophonic signals has been studied in, e.g., (Bauer, 1961; Blauert, 1997). Blauert's term
\summing localization" corresponds to manipulation of level and/or time delay dierences in stereophonic reproduction in such a way that the sound image
appears shifted from the position between the loudspeakers ((Blauert, 1997, pp.
201-271)). Damaske studied loudspeaker reproduction issues and formulated the
basic theories further in the TRADIS project (True Reproduction of All Directional Information by Stereophony) (Damaske, 1971). He conducted studies
on sound image quality deterioration as a function of listener placement. The
transaural theory was rened and to some extent revitalized by works of Cooper
and Bauck (Cooper and Bauck, 1989). They created a concept of spectral stereo
and gave a new term to crosstalk canceled binaural presentation, transaural processing. The transaural stereo concept originally applied simplied head models
for crosstalk cancelling. These techniques have been further developed by, for
example, (Kotorynski, 1990; MacCabe and Furlong, 1991; Rasmussen and Juhl,
1993; Juhl, 1993; Walsh and Furlong, 1995) to include improved head models and
more sophisticated signal processing techniques. Recently, adaptive processing
systems have been presented by (Nelson et al., 1992, 1995, 1996a,b) that take into
account multiple point equalization and the possibility of a wider listening area.
139

140

6. Crosstalk Canceled Binaural Reproduction

Also, the concept of stereo dipole has been introduced, where closely spaced loudspeakers are used to generate virtual sources (Bauck and Cooper, 1996; Watanabe
et al., 1996; Takeuchi et al., 1997; Kirkeby et al., 1997). A method for robustness
analysis of crosstalk canceling using dierent loudspeaker spacings has been proposed by Ward and Elko (1998, 1999). Fundamental and theoretical limitations
of cross-talk canceled binaural audio processing have been studied by Kyriakakis
(1998); Kyriakakis et al. (1999). An excellent recent study of cross-talk canceling
methods has been published by Gardner (1998a).
In this chapter, methods for crosstalk canceling and virtual loudspeaker implementation in symmetrical and asymmetrical cases will be introduced. Filter
design and practical implementation issues are discussed. As a case study, a novel
lter design algorithm for virtual loudspeakers based on joint minimum-phase reconstruction and warped lters is presented. The quality of the lter design has
been veried in a round robin subjective test on virtual home theatre algorithms
(Zacharov et al., 1999).
6.1

Theory of Crosstalk Canceling

The listening process in a two-channel loudspeaker case can be formulated in matrix notation. The basic theory overviewed here originates to works by (Schroeder
and Atal, 1963), and (Cooper and Bauck, 1989), the notation follows that of
(Gardner, 1995). A situation is considered, which is depicted in Fig. 6.1, where
x^l (n) and x^r (n) are binaural signals delivered to the speakers, and yl (n) and yr (n)
are the resulting signals registered at the listener's ears. The sound propagation
in a stereophonic system can be described by the following equations:
y (n) = H (z )^x(n)
(6.1)
where






y
x
^
H
l (n)
l (n)
ll (z ) Hrl (z )
y (n) = y (n) ; x^(n) = x^ (n) ; H (z ) = H (z ) H (z )
(6.2)
r
r
lr
rr
If x(n) = [xl (n)xr (n)]T is the binaural signal that is to be delivered to the ears,
an inverse matrix G(z) must be found to the system transfer matrix H (z) such
that G(z) = H (z) 1 and x^(n) = G(z)x(n). The exact inverse matrix can be
analytically written as


1
H
H
rr (z )
rl (z )
G(z ) =
(6.3)
Hlr (z ) Hll (z )
Hll (z )Hrr (z ) Hlr (z )Hrl (z )
However, an ideal inverse lter can not necessarily be computed (due to possible non-minimum phase functions in the denominator) analytically, and thus the

6.1

141

Theory of Crosstalk Canceling

x
l

H
H

lr

rl

ll

y
l

rr

Figure 6.1: Loudspeaker-to-ear transfer functions in stereophonic listening arrangement.

equality y(n) = H (z)^x(z) = H (z)G(z)x(n) should be expressed in the form


yl (n)
yr (n)

Hll (z ) Hrl (z )
Hlr (z ) Hrr (z )



Gll (z ) Grl (z )
Glr (z ) Grr (z )



xl (n)
xr (n)

(6.4)

From this point forward, the formulations are divided into two categories: symmetrical and asymmetrical listening position.
6.1.1 Symmetrical Crosstalk Canceling
In the rst case, symmetry is assumed in the listening situation and head geometry (that is, Hlr (z) = Hrl(z) and Hll(z) = Hrr (z)). Thus Eqs. 6.1, 6.2 and 6.4
can be simplied using ipsilateral responses (Hi(z) = Hll(z) = Hrr (z)) for the
HRTF to the same side and contralateral responses (Hc(z) = Hlr (z) = Hrl(z) for
the HRTF to the opposite side:

 


yl (n) = Hi (z ) Hc(z )
x^l (n)
(6.5)
yr (n)
Hc (z ) Hi (z )
x^r (n)


yl (n)
yr (n)

Gi (z ) Gc(z )
Gc (z ) Gi (z )



Hi (z ) Hc(z )
Hc(z ) Hi (z )



xl (n)
xr (n)

(6.6)

142

6. Crosstalk Canceled Binaural Reproduction

- +

1
i

+ H

- H

1
H

x
l

Figure 6.2: Shuer implementation of cross-talk canceling lters in a symmetric

listening arrangement.
The exact inverse lter can similarly be written as


1
H
H
i (z )
c (z )
G(z ) = 2
Hc (z ) Hi (z )
H (z ) H 2 (z )
G

(6.7)

(Cooper and Bauck, 1989) have proposed a \shuer" structure for the realization
of crosstalk canceling lters. This system involves generation of the sum and
dierence for input signals xl (n) and xr (n), and undoing the sum and dierence
after the ltering, as depicted in Fig. 6.2. The sum and dierence operations are
made possible by a unitary matrix , which is called the shuer matrix or MS
matrix:


1
1
1
D=p 1 1
(6.8)
2
D

It can be shown that this shuer matrix diagonalizes the matrix G(z) using a
similarity transformation:
#
"
1
0
(6.9)
D 1 G(z )D = 0Hi (z)+Hc (z)
1
Hi (z ) Hc (z )
The shuer structure is illustrated in 6.2. The normalizing gains can be commuted to a single gain of 1/2 for each channel, or just be ignored.
In this work, three dierent crosstalk cancellation approaches were considered.
In (Cooper and Bauck, 1989), it was suggested that the use of a computational
model (diraction around a rigid sphere) would result in generalized transfer
functions. These lters (Cooper approximations), although not as accurate, were
claimed to give a wider and more stable listening area than with individual or
dummy-head based crosstalk cancelers. In (Gardner, 1995), the following simple
formulas for Hi(z) and Hc(z) were used1 .
Hi (z ) = 1; Hc (z ) = gz m Hlp(z )
(6.10)
D

Gardner (1998a) presents more advanced methods of cross-talk canceler design, but these
were not applied in the current work.

6.1

143

Theory of Crosstalk Canceling

Magnitude responses of 30 cross-talk canceling filters 1/(Hi(z)+Hc(z))

Magnitude (dB)

20
Measured HRTF, subject: jyhu
Minimum-phase reconstruction
Gardner approximation
Cooper approximation
0

-20

10

10

Frequency (Hz)
Magnitude responses of 30 cross-talk canceling filters 1/(Hi(z)-Hc(z))

Magnitude (dB)

20
Measured HRTF, subject: jyhu
Minimum-phase reconstruction
Gardner approximation
Cooper approximation
0

-20

10

10
Frequency (Hz)

Figure 6.3: Magnitude responses of three crosstalk canceling lters for sources at
30o azimuth.
where g < 1 is the interaural gain, m is the approximated frequency-independent
ITD in samples, and Hlp(z) is a one-pole lowpass lter that models the frequencydependent head shadowing:
1 a
(6.11)
Hlp(z ) =
1 az 1
where coecient a determines the lowpass cuto.
In Figs. 6.3 and 6.4, results for modeling crosstalk canceling with the above
discussed two methods are presented. These transfer functions are compared
to crosstalk canceling lters based on measured HRTFs (Riederer, 1998b). The
loudspeakers in the simulation were placed at 30o azimuth. From the magnitude
response plots of Fig. 6.3 it is clearly seen that the two model-based approaches
(Cooper and Gardner approximation) give predictable results only to approximately 1-2 kHz. According to Cooper and Bauck (1989, 1992), the crosstalk
canceling lters, although not minimum-phase, are of joint minimum phase, that
is, they have a common excess phase which is close to a frequency-independent
delay. The delay-normalized crosstalk canceling lters are then minimum-phase.
Thus the shuing lters may be dened by their magnitude only, and the phase
may be calculated, e.g., using minimum-phase reconstruction. This statement
is veried in Fig. 6.4, where the measured HRTF-based crosstalk canceler and

144

6. Crosstalk Canceled Binaural Reproduction

Group delay of 30 cross-talk canceling filters 1/(Hi(z)+Hc(z))

Group delay (ms)

3
2
1

Measured HRTF, subject: jyhu

Minimum-phase reconstruction
Gardner approximation
Cooper approximation

0
10

10

Frequency (Hz)
Group delay of 30 cross-talk canceling filters 1/(Hi(z)-Hc(z))
Group delay (ms)

3
2
1

Measured HRTF, subject: jyhu

Minimum-phase reconstruction
Gardner approximation
Cooper approximation

0
10

10
Frequency (Hz)

Figure 6.4: Group delay of three crosstalk canceling lters for sources at 30o
azimuth.
the minimum-phase reconstructed version dier in group delay up to 5-6 kHz
approximately by a frequency-independent bulk delay. According to Kotorynski (1990), the phase distortion at higher frequencies caused by minimum-phase
approximation should not be neglected, but separately modeled using an allpass
lter.

6.1.2 Asymmetrical Crosstalk Canceling

In many cases the listening situation is not symmetrical. Such examples are the
interior of a car, listening to music in a concert hall, or at home on a couch
which is quite not symmetrically placed between the speakers. The asymmetrical
listening condition leads to the fact that the shuer structures presented in the
previous section are not anymore valid. It is possible to resort to use a crosstalk
canceling system illustrated in Fig. 6.1 which uses four lters. The exact inverse
matrix G is given by Eq. 6.3. The problem of asymmetrical crosstalk cancellation
and virtual source synthesis is discussed in Section 6.2.3.

6.2

Virtual Source Synthesis

145

6.1.3 Other Crosstalk Canceling Structures

The methodology for cross-talk canceling introduced by Schroeder and Atal
(1963) and the shuer structure proposed by Cooper and Bauck (1989) have
been the basis for the current research. There are, however, many other crosstalk
canceling structures that have been proposed in the literature. An excellent
overview of existing methods is presented in (Gardner, 1998a), including feedforward and recursive crosstalk canceling for both symmetric and asymmetric
listening positions.
6.2

Virtual Source Synthesis

An interesting application of cross-talk canceled binaural technology is the concept of virtual sources, rst discussed in (Cooper and Bauck, 1989). The term virtual loudspeaker (VL, or virtual speaker, VS) and VL synthesis has been invented
to mean the generation of virtual sources using cross-talk canceling techniques.
This method consists of two stages: a) implementation of binaural synthesis for
mono- or stereophonic source signals, and b) crosstalk cancellation for presentation with two loudspeakers. These steps may be combined to single ltering
task in order to optimize computational eciency. A topic related to virtual
source synthesis is stereo widening, which is a general term representing methods that enhance or exaggerate the stereophonic image of sounds in two-speaker
reproduction. This topic is discussed in Section 6.3.
6.2.1 Symmetrical Listening Position
In the previous section the treatment was restricted only to the problem of
crosstalk canceling. In many applications of interest it is, however, desired to synthesize a virtual source or multiple virtual sources for loudspeaker listening. In
other words, virtual source synthesis combines binaural processing and crosstalk
cancellation. A schematic for signal transmission in virtual source synthesis is
illustrated in Fig. 6.5. Signals xl (n) and xr (n) are raw input signals that are processed with binaural lters H 0 and cross-talk canceling lters G to deliver signals
yl (n) and yr (n) at the listener's ears. In other words, the lattice structure of Fig.
6.2 can be expanded to include the transfer functions Hi0(z) and Hc0 (z) that, for
example, place the virtual source at +90o azimuth. It may be assumed that the
loudspeakers are placed at 10o azimuth relative to the listener, and thus those
HRTFs will be used for crosstalk canceling transfer functions Hi(z) and Hc(z).
This structure, as well as all the other shuer structures, can be realized using
two digital lters.
The structure depicted in Fig. 6.6 can also be computed using a direct approach without the lattice form. In this case the transfer functions of the mono-

146

6. Crosstalk Canceled Binaural Reproduction

x
l

B in a u ra l S y n th e s is
'

G
x

C ro s s -ta lk C a n c e lla tio n

x
l

H
H

H
c

H
i

Figure 6.5: Signal transmission for cross-talk cancellation in virtual loudspeaker

synthesis.
x

H
i

'

'

H
l

+ H
+ H

- H
- H

'
c

'
c

x
l

Figure 6.6: Shuer structure for VL synthesis of one virtual source.

phonic-to-crosstalk canceled converter are of the form:
H (z )Hi0 (z ) Hc (z )Hc0 (z )
x^l (n) = i
xl (n)
Hi2 (z ) Hc2 (z )
x^r (n) =

Hi (z )Hc0 (z ) Hc (z )Hi0 (z )
xr (n)
Hi2 (z ) Hc2 (z )

(6.12)
(6.13)

6.2

147

Virtual Source Synthesis

xr
xr
xl

B
A

xl

Figure 6.7: Spectral stereo shuer for synthesis of one virtual loudspeaker.
These equations will also be used subsequently in the case of asymmetrical
crosstalk cancellation. Cooper and Bauck's spectral stereo system (Cooper and
Bauck, 1989; Rasmussen and Juhl, 1993; Walsh and Furlong, 1995) uses a rearrangement of equations 6.12 and 6.13 in the following way. Two lters, A
and B, are dened which satisfy the relations x^r (n) = B (z)xr (n) and x^l (n) =
B (z )A(z )xl (n) when
H (z )Hc0 (z ) Hc (z )Hi0 (z )
B (z ) = i
(6.14)
Hi2(z ) Hc2 (z )
A(z ) =

Hi (z )Hi0 (z ) Hc(z )Hi0 (z )

Hi (z )Hc0 (z ) Hc (z )Hi0 (z )

(6.15)

This structure is illustrated in Fig. 6.7. According to (Walsh and Furlong, 1995),
this rearrangement of lters enables us to divide the system to sound localization
part (lter A) and virtual sound image equalization part (lter B). The validity of
this statement is yet to be veried although it is clear that lter B cannot contain
any interaural phase information (because it is applied to both channels), and
furthermore, lter A is only applied to one of the channels. The drawback in this
structure is that it can not be rearranged for two-channel virtual loudspeaker
implementation with two lters, as is done in the following.
The structures presented in Eqs. 6.12 and 6.13 can be directly applied to the
implementation of two virtual sources. However, if it is assumed that the listening
position and loudspeaker placements are symmetrical with respect to the median
plane, the structures can be further simplied. It is possible to reduce the order
of the lters to two with the aid of the shuer structure of Fig. 6.6. The resulting
two-lter structure for two virtual sources is illustrated in Fig. 6.8 (Kotorynski,
1990; Toshiyuki et al., 1994; Jot et al., 1995). This proposed two-lter structure
is suitable for practical implementations due to ecient realization. In matrix
form, the normalized underlying equations can be written using Eqs. 6.5, 6.6 and
6.9:
#

 " Hi0 (z )+Hc0 (z )

0
1
1
1
1
1
H
(z )+Hc (z )
0
i
GH =
(6.16)
Hi0 (z ) Hc0 (z )
1
1
4 1 1 = 0
Hi (z ) Hc (z )

148

6. Crosstalk Canceled Binaural Reproduction

H
i

'

'

- +

+ H
+ H

'

- H
- H

'

Figure 6.8: Shuer structure for VL synthesis of two sound sources.

+
x

H
i

'

+ H
+ H

'

c
c

-1

- +

'

H
H

i
i

- H
- H

'

x
l

Figure 6.9: Monophonic decorrelation shuer for synthesizing two VLs.

6.2.2 Decorrelation of Virtual Sources

In certain applications it is of interest to create two virtual sources of a monophonic input (for example, a monophonic surround channel). It is clear that
creating two virtual speakers for a monophonic input results in a sound that is
heard in the middle or even inside the listener's head. Thus it is necessary to
decorrelate the monophonic input before it is fed to the virtual speaker lters.
The most straightforward method for decorrelating the input is to invert the
phase of the other input channel as shown in Figs. 6.9 and 6.10. Other possible
methods are delaying the other channel, or performing more demanding decorrelation analysis. The principle of converting a monophonic signal to two virtual
speakers is illustrated in Fig. 6.9 (Toshiyuki et al., 1994). Informal listening tests
carried out using this structure have shown that although negation has an eect
on the tonal quality of the sound, the virtual speaker image is still audible and
quite substantial. The structure presented in Fig. 6.9 can be further simplied
by noticing that the upper lter block actually vanishes due to the decorrelative
negation. Thus, a one-lter conguration for generating two virtual speakers
from a monophonic input is possible. This structure is depicted in Fig. 6.10.
Finally, it may be concluded that simple phase inverting and structures shown
in Figs. 6.9{6.10 are preferable for virtual speaker synthesis of a monophonic
Dolby Pro Logic surround channel.

6.2

149

Virtual Source Synthesis

6.2.3 Asymmetrical Listening Position

In the case of an asymmetrical listening position, the shuing structures can not
generally be used. In Fig. 6.11, a direct implementation of two virtual loudspeakers is presented in an asymmetrical listening position. This structure requires four
parallel lters. The asymmetrical positions can, however, be taken into account
in the lter design and optimization techniques for shuer structures. The following equations (based on Fig. 6.11) can be used to create virtual sources described
by lters Hi0(z) and Hc0 (z) for an aymmetrical listening position.
H (z )Hc0 (z ) Hrl (z )Hi0 (z )
H (z )Hi0 (z ) Hrl (z )Hc0 (z )
x^l (n) = rr
xl (n) + rr
x (n)
Hll (z )Hrr (z ) Hlr (z )Hrl (z )
Hll (z )Hrr (z ) Hlr (z )Hrl (z ) r
(6.17)
x^r (n) =

H (z )Hi0 (z ) Hlr (z )Hc0 (z )

Hll (z )Hc0 (z ) Hlr (z )Hi0 (z )
xr (n) + ll
x (n)
Hll (z )Hrr (z ) Hlr (z )Hrl (z )
Hll (z )Hrr (z ) Hlr (z )Hrl (z ) l

(6.18)

6.2.4 Binaural and Crosstalk Canceled Binaural Conversion Structures

In Fig. 6.12, digital lter and shuer structures (proposed by Cooper and Bauck)
for converting binaural and crosstalk canceled binaural signals are presented (Jot
et al., 1995). The use of shuer structures is valid for symmetrical loudspeaker
and listening arrangement, as discussed in Section 6.2.1.
6.2.5 Virtual Center Channel
In certain applications such as playback of 5.1-channel audio material using two
loudspeakers, it may be desired to synthesize not only the surround channels,
x

x
m

H
i

'

- H
- H
c

x
'

-1

Figure 6.10: Monophonic decorrelation shuer for synthesizing two VLs implemented with a single lter.

150

6. Crosstalk Canceled Binaural Reproduction

H
H

ll

ll

'
c

- H
- H

rr

lr

lr

'
i

rl

H
H

- H
- H

rr

ll

'
c

ll

H
x

rr

ll

'
i

- H
- H

rr

rl

lr

lr

'

rl

'
c

lr

rl

rr

ll

'

- H
- H

rr

rl
lr

'
rl

Figure 6.11: Direct structure for VL synthesis of two sound sources in asymmetrical listening.
(c )

(a )

H
m o n o

y
'

x
r

b in a u ra l

H
c

'

H
i

1
+ H

x
l

+
c

+ H
c

c ro s s ta lk
c a n c e le d

+
-

x
r

c ro s s ta lk
c a n c e le d

b in a u ra l

'

+ H

H
i

'

- H
c

- H

'

(d )

+
r

'
i

s te re o

(b )

H
i

1
- H
c

x
l

+
r

H
i

+ H

+
c

c ro s s ta lk
c a n c e le d

x
l

b in a u ra l

H
i

- H
c

Figure 6.12: Binaural and crosstalk canceled binaural implementation and conversion structures. a) monophonic-to-binaural, b) binaural-to-crosstalk canceled,
c) stereophonic-to-crosstalk canceled, and d) crosstalk canceled-to-binaural shuf er structures (Jot et al., 1995).

x
l

6.3

Stereophonic Image Widening

151

but also the center channel using virtual loudspeaker technology. Although the
monophonic center channel may be reproduced using left and right speakers,
virtual technology may be useful in stabilizing the front image both in terms of
tonal color and listener position.
6.3

Stereophonic Image Widening

It is clear that HRTF-based virtual sound source processing aects the sound
quality in many ways. Crosstalk canceling and virtual source synthesis are aimed
at specialized applications and require a priori knowledge of the type of source
signal that is fed into the system. In the case of crosstalk canceling, the input
material is generally binaural, and when virtual source synthesis is performed,
the inputs are multiple monophonic channels.
Methods that enhance the stereophonic image (stereo widening algorithms),
on the other hand, are aimed at processing any input material, monophonic or
stereophonic. An overview of articial stereophonic image enhancement techniques is given in (Maher et al., 1996). In the following, traditional and novel
methods for stereophonic image enhancement are presented.
6.3.1 Traditional Stereophonic Image Enhancement Techniques
The traditional methods for enhancement of stereo image involve processing the
sum and dierence signals of the inputs. It is assumed that the left and right
input signals xl (n) and xr (n) consist of a monophonic portion xm (n) common to
both inputs, and of signals xl0 (n) and xr0(n) that are emanating from left and
right channels only:
xl (n) = xm (n) + xl0 (n)
(6.19)
xr (n) = xm (n) + xr0 (n)
The sum and dierence signals can be calculated from Eq. 6.19 in the following
way:
xl (n) xr (n) = xl0 (n) xr0 (n)
(6.20)
xl (n) + xr (n) = 2xm (n) + xl0 (n) + xr0 (n)
From this equation it can be seen that adding xl (n) xr (n) to xl (n) produces
xm (n) + 2xl0 (n) xr0 (n), which clearly boosts the left-only signal. Similarly,
when xl (n) xr (n) is subtracted from xr (n), the right-only signals are enhanced.
In this work, three algorithms have been investigated which take advantage of
the theory presented above.

152

6. Crosstalk Canceled Binaural Reproduction

xl

xl +xr

Dynamic
EQ

xl

Spectrum
Analyzer
+
-

MIXER
VCA

Dynamic
EQ

xl - xr

xr

Fixed
EQ

xr
CONTROL

Figure 6.13: Stereo widening algorithm presented in (Klayman, 1988).

xl

xl +xr
MIXER

xl

EQ
+

xl - xr

MIXER

xr

VCA
HPF

Delay

xr

EQ

CONTROL

Figure 6.14: Stereo widening algorithm presented in (Desper, 1995).

In (Klayman, 1988), the sum and dierence signals of the stereophonic input
are processed as illustrated in Fig. 6.13. The enhanced (not binaural in a strict
sense) output signals x^l (n) and x^r (n) are constructed by adding the unprocessed
input signals and the processed sum and dierence signals in the mixing stage.
Adaptive frequency-dependent equalization of the xl (n)+ xr (n) and xl (n) xr (n)
signals is carried out to correct for tone coloring.
Another similar algorithm is presented in the patent by (Desper, 1995). The
structure is illustrated in Fig. 6.14. The main dierence compared to the previous
design is the fact that only the dierence signal xl (n) xr (n) of the stereophonic
input is dynamically controlled and processed, the sum signal xl (n) + xr (n) is
mixed into the enhanced output without dynamic processing.
The technique presented in (Maher et al., 1996) is also based on processing of
the dierence signals of the stereophonic input as in the two previous examples.
This structure is shown in Fig. 6.15. In addition to previous algorithms, the use
of two adaptive FIR lters Hl (z) and Hr (z) is suggested. An adaptive leastmean square (LMS) algorithm (Widrow and Stearns, 1985) is used to minimize

6.3

153

Stereophonic Image Widening

Hl

xl

xl , p
+ d = xl - xr
-

l = d xl , p

r = d xr ,p

xr

xr , p
Hr

Figure 6.15: Stereo widening algorithm presented in (Maher et al., 1996).

the error terms r and l and between the desired output (xl (n)
processed outputs xl;p(n) and xr;p(n).

xr (n)) and the

6.3.2 Stereo Widening Based on 3-D Sound Processing

When 3-D sound processing algorithms are used for stereo widening, the tone
quality of the reproduced stereophonic sound becomes a more and more critical
issue. In the literature, a system that performs stereophonic image enhancement
with \placement lters" is described (Lowe et al., 1995). In this system (illustrated in Fig. 6.16) it is assumed that the inputs xl0 (n) and xr0(n) have been
processed so that the monophonic portion xm (n) common to both channels has
been extracted. The algorithm resembles a virtual loudspeaker synthesis scheme
of Fig. 6.8 with an additional delay unit for mixing the unprocessed left and
right signals with the output. A novel algorithm for stereo widening based on
crosstalk canceled HRTF processing was presented in (Valio and Huopaniemi,
1996). The proposed structure enhances the stereophonic image using virtual
speaker processing, but retains the original tone color for monophonic sounds.
This is achieved by extracting a sum signal of the inputs and applying a tone
correction lter to that portion of sound while the left and right channels are
separately processed as in virtual speaker reproduction. The system has been
further modied and optimized for computational requirements. A schematic of
the new system (modication of (Valio and Huopaniemi, 1996)) is illustrated in
Fig. 6.17. In this structure the crosstalk canceling lters Hi(z) and Hc(z) and the
virtual speaker placement lters Hi0(z) and Hc0 (z) are calculated using methods

154

6. Crosstalk Canceled Binaural Reproduction

delay

xl 0

xl , p
Left
placement
filter

+
-

Right
placement
filter

xr , p

xr 0

delay

Figure 6.16: Stereo widening algorithm presented in (Lowe et al., 1995).

g

- +

H
i

'

H
2

g
l

'
i

+ H
+ H

- H
- H

'

'

x
l

Figure 6.17: Stereo widening algorithm based on virtual loudspeaker positioning

and control of the monophonic sound.
presented in Section 6.2. The lter Hm(z) is used to compensate for coloration
and delay caused by virtual speaker processing. By choosing an appropriate distribution for the gains g1 and g2 for the pseudo-monophonic sum-signal the tonal
quality of the overall enhanced stereo image can be controlled.
6.4

Virtual Loudspeaker Filter Design

There are two major diculties in the design and implementation of virtual
loudspeaker or crosstalk canceling structures:
 limited listening area (\sweet spot"), where the loudspeaker binaural system

6.4

155

Virtual Loudspeaker Filter Design

Research Group

MacCabe and Furlong, 1991

Design Type

Filter Or-

Study

der

CT-canceled
20
Empirical?
binaural / IIR
Cooper and Bauck, 1993
CT-canceled
12
Empirical?
binaural / IIR
Kotorynski, 1995
CT-canceled
64
Empirical?
binaural / FIR
Kotorynski, 1995
CT-canceled
32
Empirical?
binaural / IIR
Jot et al., 1995
CT-canceled
12
Empirical?
binaural / IIR
Gardner, 1998
CT-canceled
8
Empirical
binaural / IIR
Table 6.1: crosstalk canceled binaural HRTF lter design data from the literature.
functions well
 undesired coloration caused by VL or crosstalk ltering
The concept of joint minimum phase in crosstalk canceling and VL design (discussed in Section 6.1) helps us to understand the general problematics of crosstalk
canceled binaural reproduction. The phase information in shuer lter structures
may be considered redundant, i.e., a minimum-phase component for the sum and
dierence signals can be reconstructed using the Hilbert transform (Oppenheim
and Schafer, 1989). Although this assumption is only valid in symmetrical listening and may cause sound image degradation (Kotorynski, 1990), it will help
in the generalization of VL lters to work on a larger listening area with less
coloration.
The task of crosstalk canceling binaural lter design diers somewhat from
binaural lter design issues that were discussed in the previous Chapter. Crosstalk
canceling and the problem of limited listening area introduce complications and
constraints that require advanced lter design techniques. In Table 6.1, results
from studies published in the literature are listed. As can be seen, crosstalk canceled binaural lter design issues have not been widely explored. By nature, the
form of crosstalk canceling lters is recursive, because the canceling process is innite. This would suggest that recursive pole-zero models can be used for modeling
crosstalk or virtual loudspeaker synthesis. The design of joint minimum-phase
cross-talk canceling lters can be carried out using similar techniques as in binaural lter design. In the following, methods for crosstalk canceled binaural lter
design are presented. The techniques are equally applicable to the design of both
crosstalk lters, and combined crosstalk and virtual loudspeaker lters.

156

6. Crosstalk Canceled Binaural Reproduction

6.4.1 Finite Impulse-Response Methods

The windowing method gives an optimal t in the least-squares sense to the
given frequency response. In VL design, however, the problem of solving multiple
transfer functions simultaneously, as is the case in the shuer structures, arises.
An optimal solution to the problem in the least-squares sense can be found by
solving for all the lters in the inverse matrix formulation simultaneously so that
the estimation errors are matched and optimized. This type of approach using
adaptive LS optimization has been taken by (Nelson et al., 1992, 1996a,b; Kirkeby
et al., 1998; Kirkeby and Nelson, 1998).
6.4.2 Innite Impulse-Response Methods
The principle of warped lter design discussed in Chapter 5 can be extended to
crosstalk canceled binaural processing. When the shuer lters can be specied
as minimum-phase systems, warped structures are an attractive solution for efcient auditory-based lter design and implementation. An outline for a warped
shuer lter (WSF) design method is as follows:
 Compute the shuer transfer functions using Eq. 6.16
 Extract the joint minimum-phase part of the transfer functions
 Warp the frequency axis using Eq. 5.19
 Design the lters using a traditional IIR lter design method (e.g., Prony's
method (Parks and Burrus, 1987)
 Unwarp the designed warped lter to second-order section implementation
using Eqs. 5.39{5.40.
In Fig. 6.18, results for designing a minimum-phase VL shuer lter for 10 90
are presented. It can be seen that a WIIR of order 15 is capable of retaining most
of the spectral features up to 15 kHz with an enhanced low frequency t. The
validity and performance of the WSF lters has been veried in a round-robin
experiment of virtual home theater (VHT) systems, discussed in Section 6.6.
6.4.3 Conclusion and Discussion
Filter design for crosstalk canceled binaural systems demands special care when
specifying the target response and the goal of optimization. Basically, the crosstalk
canceling or virtual loudspeaker lters should satisfy two goals:
 optimization for a large listening area versus the stability of the image
 optimization for a generalized set of lters that minimize sound coloration

6.5

157

System Analysis

Modeling of VL design 10-90 using minimum-phase structures, left shuffler filter

Minimum-phase Original
FIR, window: boxcar, order: 60
WFIR, lambda=0.65, window: boxcar, order: 20
IIR, design: Prony, order: 30
WIIR, lambda= 0.65, design: Prony, order: 15
WIIR2, lambda= 0.7313, design: Prony, order: 30

Relative Magnitude (dB)

20

-20

-40

-60
2

10

10
Frequency (Hz)

Figure 6.18: Virtual loudspeaker lter design for 10

WFIR, and WIIR lters.

10

90 using FIR, IIR,

To meet these specications, experiments have been carried out in listening area
widening and dierent lter design methods. As a conclusion, according to informal listening tests, FIR lters of order 40-50 and WIIR lters of order 15-20
based on the minimum-phase shuer structure proposed by Cooper and Bauck
(1989) would suce in realization of crosstalk canceling and virtual loudspeaker
implementation.
6.5

System Analysis

6.5.1 Widening the Listening Area

In this section, techniques for listening area widening and the control of tone
coloration are discussed. The \sweet spot" for listening remains a severe problem in loudspeaker binaural synthesis of audio signals. There are, however, several approaches to study and partly solve this problem. Intuitively, two-channel
stereophony is very limited in giving possibility to listening area widening due to
the precedence eect (Yost and Gourevitch, 1987). A related topic is the stability
and coloration of the image, i.e. the virtual sound image quality deterioration in
an asymmetrical listening position.

158

6. Crosstalk Canceled Binaural Reproduction

Head-Tracked Systems

One of the most prominent techniques for expanding the sweet spot is to use
tracking of the listener position (Gardner, 1998a; Kyriakakis and Holman, 1998).
This partly overcomes the problems of robust crosstalk canceling due to the fact
that the canceling lters can be steered according to listener movements. Methods
based on magnetic tracking (Gardner, 1998a) or visual tracking (Kyriakakis and
Holman, 1998) of the listener's position have been proposed. This technique
is suitable for one listener only, which is also true for most loudspeaker based
binaural reproduction systems.
Stereo Dipole Technique

A principle called \stereo dipole" has been recently introduced that uses two
closely spaced loudspeakers for virtual source synthesis in loudspeaker listening (Watanabe et al., 1996; Takeuchi et al., 1997; Kirkeby et al., 1998; Kirkeby
and Nelson, 1998). Based on theoretical formulations, the conclusions were that
bringing the loudspeakers closer to each other enhances and stabilizes the virtual speaker image, and widens the listening area. This has also been veried
by robustness analysis of crosstalk canceling for dierent loudspeaker placings
(Ward and Elko, 1998, 1999). One of the inherent problems of the stereo dipole
technique, however, is the computational complexity that results from the fast deconvolution and frequency-dependent regularization (Kirkeby and Nelson, 1998)
algorithms. This is due to the goal of optimal inversion of the two-by-two system
of two-loudspeaker listening, not taking into account the joint minimum-phase
properties of the system. Warped FIR structures have been proposed for the
stereo dipole technique (Kirkeby et al., 1999) to enhance computational eciency,
resulting in a two-by-two matrix of WFIR lters each containing 32 coecients.
This presents a clear reduction in computation, but is still outside of the scope
of ecient real-time systems (such as the WSF method proposed by the author
in Section 6.4.2).
Presenting Binaural Material over Multiple Loudspeakers

In a method proposed by (Bauck and Cooper, 1992, 1996), the center image can
be stabilized by adding a third loudspeaker, a center channel. This addition also
introduces natural widening of the listening area by adding a second \sweet spot"
to the listening area. An exact solution can be found, which uses four lters in
realization, but the other listener gets a reversed image. According to (Bauck
and Cooper, 1996), a nonreversed image can also be computed, but the result is
not exact due to an overdetermined solution.
It is also possible to use pairwise binaural presentation with crosstalk cancellation in multiple speaker systems. For example, using two frontal and two rear
speakers with pairwise crosstalk cancellation it is possible to create convincing

6.5

System Analysis

159

front-back discrimination and a more stable 360 panning of virtual sources. Furthermore, using intermediate formats such as Ambisonics (Malham and Myatt,
1995) for binaural rendering2 it is possible to apply spatial transcoding of signals
for dierent reproduction formats, and to reduce the cost of binaural synthesis
of multiple sources.
6.5.2 Analysis of Crosstalk Canceled Binaural Designs
In this experiment, dierent methods for crosstalk canceled binaural lter design
were compared. The subjective and objective quality of two crosstalk canceling
methods based on computational models was compared to HRTF-based VL and
crosstalk designs. The computational models were based on Rayleigh's diraction
formula, and the approximation proposed by Gardner (Eqs. 6.10 and 6.11). The
HRTF database consisted of measurements made on 33 human subjects (Riederer,
1998b). The design goal was to create virtual loudspeakers at 90o azimuth when
the physical speaker placement was 10o .
As discussed earlier, in the symmetrical case the crosstalk canceling shuer
structures are approximately of joint minimum phase, that is, the phase response
of Hi(z) + Hc(z) and Hi(z) Hc(z) can be calculated using minimum-phase reconstruction. Thus the transfer functions can be derived using magnitude only
approximation without phase constraints. In Fig. 6.19, crosstalk canceling transfer functions for 33 subject HRTFs are presented. A mean of the magnitude
values was also calculated, which is shown as the thick line in Fig. 6.19. Although it is clear that some of the idiosyncratic information (above 4-5 kHz)
present in HRTFs is smoothed when the mean value is taken on a frequencypoint basis (see, e.g., (Mller et al., 1995) for discussion), in this context the
use of a generalized magnitude response is motivated. It is further possible to
smooth the mean responses by, e.g., 1=3 octave or auditory smoothing in order
to remove such magnitude uctuations that are not audible according to psychoacoustic theory. In theory, also the VL positioning lters functions Hi0(z) and
Hc0 (z ) should be of nearly joint minimum phase, so the same technique could be
applied to yield minimum-phase realizations of VL shuer lters.
In Fig. 6.20, the results for VL design based on four techniques are illustrated.
The empirically based VL structures used data measured from one test person.
The solid line of Fig. 6.20 represents the VL design for the test person's HRTFs.
The dashed line in the gure represents a VL design where computed magnitude
means from the population of HRTFs have been used both for virtual loudspeaker
placement ltering (the transfer functions Hi0(z) and Hc0 (z) as depicted in Figs.
5.6 and 5.8), and for crosstalk canceling. This results in a minimum-phase system. The dash-dotted and dotted lines represent the Cooper and Gardner VL
approximations, respectively. It can be clearly seen that the computational mod2

This approach is called the \binaural B format" (Jot et al., 1998b).

160

6. Crosstalk Canceled Binaural Reproduction

Crosstalk canceling design for +-10, 1/(Hi(z)+Hc(z)), 33 subjects

Amplitude (dB)

20
All subjects
Magnitude mean
0

-20 2
10

10

10

Crosstalk canceling design for +-10, 1/(Hi(z)-Hc(z)), 33 subjects

Amplitude (dB)

20

0
All subjects
Magnitude mean
-20 2
10

10
Frequency (Hz)

10

Figure 6.19: Crosstalk canceling magnitude responses for loudspeakers placed at

10o azimuth based on 33 human subject HRTFs (shown as dotted lines). The
line marked with (+) represents a calculated mean of the magnitude responses.

els resemble the empirical results only up to approximately 2 kHz. In informal

listening tests, the VL structure based on a mean minimum-phase approximation
performed in a satisfactory manner when compared to individualized VL design.
Coloration caused by sharp peaks in the individualized VL design was reduced
(seen in Fig. 6.20), but the listening area remained stable. A minimum-phase
realization of the individualized VL lters was also tested, and no audible dierence to the non-minimum-phase version was found. The performance of VL and
crosstalk designs based on computational HRTFs, however, was found unsatisfactory. The virtual loudspeaker image was unstable and often the test signal
was perceived as emanating from one of the speakers.
These results suggest that generalization and smoothing of symmetrical crosstalk
canceling structures could reduce tone coloration and expand the listening area.
More thorough listening tests should be carried out, however, to fully support
these conclusions.

6.6

161

Subjective Evaluation of Virtual Surround Systems

VL design 10-90, left filter, shuffler structure

Magnitude (dB)

20
Test subject: jyhu
Cooper approximation
Gardner approximation
Generalized min-phase VL
0

-20
10

10

VL design 10-90, right filter, shuffler structure

Magnitude (dB)

20

Test subject: jyhu

Cooper approximation
Gardner approximation
Generalized min-phase VL

-20
10

10

Frequency (Hz)

Figure 6.20: VL design magnitude responses for loudspeakers placed at 10o

azimuth and virtual speaker placement of 90o azimuth. The shuer structure
was used.
6.6

Subjective Evaluation of Virtual Surround

Systems

In this section, overview and results of a round robin subjective evaluation experiment on virtual surround systems are provided (Zacharov et al., 1999; Zacharov
and Huopaniemi, 1999). The goal of the experiment was to compare discrete 5
channel reproduction to reproduction of 5 channel material over two loudspeakers
placed at 30 or at5 . Furthermore, the aim was to compare the quality and
performance of dierent virtual surround algorithms provided by industry and
academia (Zacharov et al., 1999).
6.6.1 Background
As the development of high quality multichannel audio encoding, storage media
and broadcast techniques evolves, the availability of multichannel audio has become widespread. While matrixed multichannel methods (e.g. Dolby Surround)
have been widely available for some time, the benets of discrete multichannel
reproduction are nding increasing favor in professional and consumer markets,
for example, with new audiovisual storage formats such as DVD.
The so-called virtual surround or virtual home theater (VHT) (Olive, 1998)

162

6. Crosstalk Canceled Binaural Reproduction

aim to faithfully reproduce the spatial sound qualities of 5.1 channel systems using
only two loudspeakers. Virtual loudspeaker technology enables the virtualization
of surround left and right channel information. Thus the surround channels
appear to emanate from virtual sources placed outside the physical stereo setup
of the two loudspeakers. Virtualization can also be created for the left and
right signals in order to, for example, expand the stereo base. Furthermore,
virtualization may be applied for the center channel, or it can be left unprocessed
(the traditional phantom image)3.
6.6.2 Experimental Procedure and Setup
This section summarizes the experimental procedure and setup for the virtual
surround experiment performed in the two sites. At each site a dierent listening
room and listening panel were employed as described below. In all other respects
the experiments performed at each site were identical. For a full description of
the design, the interested reader is referred to (Zacharov et al., 1999).
The rst of the two experiments was performed in the new Nokia Research
Center (NRC), Speech and Audio Systems Laboratory listening room, which is
fully conformant with ITU-R BS.1116-1 (ITU-R, 1994a) as illustrated in Fig. 6.21.
The second experiment was performed at the AES 16th International Conference
in a smaller meeting room. For reproduction, Genelec 1030A speakers were employed for all of the experiments consisting of a two-way speaker with a frequency
response of 52 Hz { 20 kHz (3 dB).
The setup for the 5-channel reproduction was in accordance with ITU-R
BS.775-1 (ITU-R, 1994b) as far as possible and thus speakers were placed at
the normal 0, 30, 110 angles, with the speaker's axis at average ear height.
In addition, a pair of speakers were set up at an angle of 5 to enable virtual surround reproduction using closely spaced loudspeakers. Loudspeakers were placed
behind an acoustically transparent screen to limit any visual bias eects. As only
discrete 5-channel material was employed for this experiment, no low frequency
energy (LFE) channel or loudspeaker was available.
For the experiment, a rank order procedure was chosen due to its eciency
and the simplicity of the rank ordering task when employing untrained listeners
(Zacharov et al., 1999). Furthermore, as no xed reference was employed, the
rank order procedure enabled the comparison of the perceived quality dierence
between a discrete 5-channel system and virtual home theatre systems, without
the assumption that the 5-channel is either inferior or superior in performance.
3

It should be noted that all virtual surround or virtual home theater concepts discussed in
this experiment refer to 2-loudspeaker reproduction of 5.1 audio material in such a way that at
least the surround (and possibly also the center, left and right) audio channels are processed
virtually.

6.6

Subjective Evaluation of Virtual Surround Systems

163

Figure 6.21: Layout of the new NRC ITU-R BS.1116-1 (ITU-R, 1997) compliant
listening room and loudspeaker setup (Zacharov et al., 1999).
Tested Systems

A total of six VHT systems were compared to a reference discrete 5 channel

reproduction. A complete list of the proponents and the details of the systems
can be found in Table 6.2. For the VHT experiment, a set of warped shuer
lters (WSF), described in detail in Section 6.4.2 were designed and employed
using the following specications. This set of lters designed by the author was
called PBVS (Polar Bear Virtual Surround) in the VHT experiment (Zacharov
et al., 1999; Zacharov and Huopaniemi, 1999).
 Shuer lter structure was used for 110 virtual surround channel generation.
 Two minimum-phase warped lters of order 20 were designed4
 The lters were implemented by unwarping the responses to second-order
sections.
4

Note that there is an error in (Zacharov et al., 1999). The order of the WIIR lters used
in the experiment was 20.

164

6. Crosstalk Canceled Binaural Reproduction

Proponent

System

Harman Multimedia

VMAx 3D Virtual Theatre

Aureal
ductor

Reprod- Center
uction
chanangle
nel

MIPS
@
48
kHz

System description

Virtual

A3DS

30

85 lter coecients, memory

400(word size)

Phantom

20

Virtual front channels and

decorrelation processing, coefcient memory: 150 words, delay memory: 1k words (inc.
decorrelator)

SRS Labs, Inc.

TruSurround

30

Phantom

4.7

IIR lters with 18 lter coecients / Total ROM code

+ lter coecients = 306,
RAM=36

Helsinki University of Technology

PBVS

30

Phantom

6.4

Warped IIR lters of order 20

Sensaura Ltd

Sensaura Virtual
Surround

30

Phantom

5.2

25 coecients per lter, no additional memory required

Dolby
ries

Dolby Virtual
Surround

Phantom

5

10 lters and 10 delays

Semicon-

Laborato-

Table 6.2: Summary of proponent systems, based upon the information provided
by manufacturers (Zacharov et al., 1999).
6.6.3 Test Items and Grading
Four program items were selected for the experiment. The items were selected
to provide a range of spatial sound cues and dierent types of programme.
 Blue bell railway, BBC. Scene consisting of steam training pulling away
from station and approaching bridge. Contains country atmosphere with
directional cues, and panning eects.
 Tipsy Gypsy, BBC. A concert at the Albert hall, consisting of audience
cheering, applause and the conductor talking.
 Rain storm. Sample consisting of a thunder roll followed by the sound of
hard rain.
 Felix Mendelssohn-Bartoldy, Symphony No. 4. Live recording at the Neues
Gewandhaus, Leipzig, Deutsche Telekom, 1993.
Two grading scales were employed to evaluate the perceived spatial and timbral quality of reproduction of the seven systems under test. The following questions were posed to each listener.
 Rank order the samples by spatial sound quality (1 = lowest rank, 7 =
highest rank)

6.7

Discussion and Conclusions

165

When evaluating the spatial sound quality, please consider all aspects of
spatial sound reproduction. This might include the locatedness or localisation of the sound, how enveloping it is, it's naturalness and depth.
 Rank order the samples by timbral quality (1 = lowest rank, 7 = highest

rank)
When considering aspects of the timbre quality, please consider timbre as
a measure of the tone colour. Timbre can be considered as the sensory
attribute which allows two similar sounds, of the same pitch and loudness,
to be dierent, for example a clarinet and a cello. Any audible distortions
can also be considered as an aspect of the timbral quality.

6.6.4 Results and Discussion

The results of the VHT experiment are shown in Figs. 6.22{6.23 (Zacharov and
Huopaniemi, 1999). The PBVS algorithm using WSF lters is marked \5" in
the NRC test and \25" in the AES16 test. From the results it can be seen that
the performance of the warped shuer lters is very good, above average when
compared to other tested methods (resulting second and third spatially, and rst
and second timbrally of the VHT algorithms). As a summary, the following
conclusions from the test can be made:
 The WSF structure performed very well in the test compared to commercial
algorithms and implementations. The WSF algorithm is computationally
ecient and provides good virtual source imaging with little coloration.
 Results are statistically signicant and considered meaningful.
 Results are similar between both sites.
 Discrete 5 channel is signicantly superior.
 There are strong correlations between the ranking for both spatial and
timbre quality for all programme items.
6.7

Discussion and Conclusions

In this chapter, concepts and methods for crosstalk canceled binaural reproduction were discussed. Theory and formulas for virtual source synthesis in symmetrical and asymmetrical listening setups were reviewed. The validity of joint
minimum-phase reconstruction was veried. A new virtual loudspeaker lter design method based on the shuer structure using warped lters was introduced.
The performance of the warped shuer lter was veried in subjective experiments.

166

6. Crosstalk Canceled Binaural Reproduction

Figure 6.22: Subjective experiment result from the NRC experiment (Zacharov
and Huopaniemi, 1999).

Figure 6.23: Subjective experiment result from the AES16 experiment (Zacharov
and Huopaniemi, 1999).

Chapter 7
Conclusions
In this work, issues in 3-D sound and virtual acoustics in the eld of multimedia
were treated. New concepts, methods, algorithms and experimental results were
presented, and their importance to virtual acoustics modeling was discussed. The
topic is both up-to-date and innovative from the research point of view, because
the integration of dierent media such as telecommunications, multimedia and
virtual reality has placed new requirements and standards for audio and audio
signal processing. The concept of virtual acoustics was treated, which consists
of source, environment and receiver modeling. In this work, new methods were
introduced for parametrization, lter design and real-time rendering of virtual
acoustics, allowing for more realistic 3-D audio presentation with ecient computation.
In Chapter 2, an overview of virtual acoustic technology was given. The author's contribution to the DIVA system and to MPEG-4 standardization were
presented, and overall design and implementation methodology for virtual acoustics was discussed.
In Chapter 3, topics in sound source parametrization and modeling were studied. The author has contributed to source directivity modeling by introducing
two novel techniques: a) Directional ltering, and b) Set of elementary sources.
Furthermore, the principle of reciprocity was applied to modeling the sound radiation from the human mouth. This novel technique enables fast and reliable
approximation of the directivity characteristics of human speech for virtual reality
modeling purposes.
Chapter 4 dealt with real-time room acoustical issues, concentrating on lter
design aspects for the phenomena of absorption due to air and material re ections.
The author proposed new methods for ecient low-order lter design for air
absorption and acoustical material properties.
Chapter 5 concentrated on HRTF approximation and lter design. New design
methods for binaural lters based on balanced model truncation and frequency
warping were proposed by the author. Methods for subjective and objective evaluation of binaural lter design were introduced. The aim in applying a binaural
167

168

7. Conclusions

auditory model for objective evaluation of binaural lter design quality is to be

able to predict the subjective preferences, i.e., what is relevant in binaural processing for perception. The results provide guidelines and recommended practice
for binaural lter design and are useful for both research and applications of
binaural technology.
In Chapter 6, binaural techniques were extended for loudspeaker reproduction by crosstalk canceling mechanisms and virtual loudspeaker technology. The
author introduced a warped shuer lter (WSF) design method, which enables
low-order designs of minimum-phase crosstalk cancelers and virtual loudspeakers. The performance of the WSF method was veried in a round-robin subjective
listening experiment.
The main results of this thesis can be summarized as follows:
 Sound source directivity can be incorporated in virtual reality systems using
two methods: a) Directional ltering, b) Set of elementary sources.
 A reciprocity method can successfully be applied to estimation of directional
sound radiation from the human mouth.
 Air absorption and boundary material re ection eects can be eciently
modeled using low-order digital lters
 The high-frequency spectral content present in HRTFs can be smoothed
using auditory criteria without degrading localization performance.
 A binaural auditory model can be used to give a quantitative prediction of
perceptual HRTF lter design performance.
 Auditory lter design principles using, for example, frequency warping or
balanced model truncation result in more ecient and more perceptually
relevant binaural processing.
 Virtual loudspeaker lters designed using a joint minimum-phase shuer
structure and auditory design principles provide high-quality virtual source
imaging for virtual surround applications.
The author has contributed to source modeling, environment modeling, and
listener modeling in virtual acoustics. The results presented in this thesis may
directly be applied to digital audio signal processing in general, and more specically the design and implementation of virtual acoustic environments. The theory
and applications of 3-D sound and virtual acoustics is still a rapidly evolving area,
and both commercial and scientic interest in binaural signal processing has increased. Interesting topics such as binaural models of hearing, auralization based
on auditory optimization, and spatial transcoding of audio formats remain as
future challenges for the author.

Bibliography
Abel, J. 1997. Three-dimensional virtual audio display employing reduced complexity imaging lters, U.S. patent no. 5,659,619.
Abel, J. and Foster, S. 1997. Method and apparatus for ecient presentation of
high-quality three-dimensional audio, U.S. patent no. 5,596,644.
Abel, J. and Foster, S. 1998. Method and apparatus for ecient presentation of
high-quality three-dimensional audio including ambient eects, U.S. patent no.
5,802,180.
Allen, J. B. and Berkley, D. A. 1979. Image method for eciently simulating
small-room acoustics, Journal of the Acoustical Society of America (4): 943{
950.
Asano, F., Suzuki, Y. and Sone, T. 1990. Role of spectral cues in median plane
localization, Journal of the Acoustical Society of America (1): 159{168.
Atal, B. S. and Schroeder, M. R. 1966. Apparent sound source translator, U.S.
patent no. 3,236,949.
Barron, M. 1993. Auditorium Acoustics and Architectural Design, E & FN Spon.
Bass, H. and Bauer, H.-J. 1972. Atmospheric absorption of sound: Analytical
expressions, Journal of the Acoustical Society of America (3): 821{825.
Bauck, J. L. and Cooper, D. H. 1992. Generalized transaural stereo, Presented
at the 93rd Convention of the Audio Engineering Society, preprint 3401, San
Francisco, CA, USA.
Bauck, J. L. and Cooper, D. H. 1996. Generalized transaural stereo and applications, Journal of the Audio Engineering Society (9): 683{705.
Bauer, B. B. 1961. Stereophonic earphones and binaural loudspeakers, Journal
of the Audio Engineering Society : 148{151.
Bech, S. 1993. Training of subjects for auditory experiments, Acta Acustica
: 89{99.
65

88

52

44

169

170

BIBLIOGRAPHY

Begault, D. R. 1994. 3-D Sound for Virtual Reality and Multimedia, Academic
Press, Cambridge, MA, USA, p. 293.
Beliczynski, B., Kale, I. and Cain, G. D. 1992. Approximation of FIR by IIR
digital lters: An algorithm based on balanced model reduction, IEEE Transactions on Signal Processing (31): 532{542.
Blauert, J. 1997. Spatial Hearing. The Psychophysics of Human Sound Localization, MIT Press, Cambridge, MA, USA, p. 494.
Blommer, M. A. 1996. Pole-zero Modeling and Principal Component Analysis of
Head-Related Transfer Functions, PhD thesis, University of Michigan, p. 165.
Blommer, M. A. and Wakeeld, G. H. 1994. On the design of pole-zero approximations using a logarithmic error measure, IEEE Transactions on Signal
Processing (11): 3245{3248.
Blommer, M. A. and Wakeeld, G. H. 1997. Pole-zero approximations for headrelated transfer functions using a logarithmic error criterion, IEEE Transactions on Speech and Audio Processing (3): 278{287.
Borish, J. 1984. Extension of the image model to arbitrary polyhedra, Journal
of the Acoustical Society of America (6): 1827{1836.
Botteldooren, D. 1995. Finite-dierence time-domain simulation of lowfrequency room acoustic problems, Journal of the Acoustical Society of America
(6): 3302{3308.
Brandenburg, K. and Bosi, M. 1997. Overview of MPEG audio: Current and future standards for low-bit-rate audio coding, Journal of the Audio Engineering
Society (1/2): 4{20.
Brown, C. P. and Duda, R. O. 1998. A structural model for binaural sound
synthesis, IEEE Transactions on Speech and Audio Processing : 476{488.
Brungart, D. 1998. Near-eld Auditory Localization, PhD thesis, Massachusetts
Institute of Technology, p. 253.
Brungart, D. and Rabinowitz, W. 1996. Auditory localization in the near-eld,
Proceedings of the International Conference on Auditory Display, Palo Alto,
California, USA.
Brungart, D. S., Durlach, N. I. and Rabinowitz, W. R. 1997. Three-dimensional
auditory localization of nearby sources, Journal of the Acoustical Society of
America
(5): 3140.
40

42

75

98

45

102

171

BIBLIOGRAPHY

Calamia, P. and Hixson, E. 1997. Measurement of the head-related transfer function at close range, Journal of the Acoustical Society of America (5): 3117.
Chen, B.-S., Peng, S.-C. and Chiou, B.-W. 1992. IIR lter design via optimal
Hankel norm approximation, IEE Proc. G: Circuits, Devices Syst. (5): 586{
590.
Chen, J., Veen, B. V. and Hecox, K. E. 1995. A spatial feature extraction and regularization model for the head-related transfer function, Journal of the Acoustical Society of America (1): 439{452.
Chung, J. Y. and Blaser, D. A. 1980. Transfer function method of measuring the
in-duct acoustic properties. Part I: Theory, and Part II: Experiment, Journal
of the Acoustical Society of America.
Cook, P. R. and Trueman, D. 1998. A database of measured musical instrument
body radiation impulse responses, and computer applicatoins for exploring and
utilizing the measured lter functions, Proceeding of the International Symposium on Musical Acoustics, Leavenworth, WA, USA.
Cooper, D. H. 1982. Calculator program for head-related transfer function, Journal of the Audio Engineering Society (1/2): 34{38.
Cooper, D. H. and Bauck, J. L. 1989. Prospects for transaural recording, Journal
of the Audio Engineering Society (1/2): 3{19.
Cooper, D. H. and Bauck, J. L. 1992. Head diraction compensated stereo system,
U.S. patent no. 5,136,651.
Dalenback, B.-I. 1995. A New Model for Room Acoustic Prediction and Auralization, PhD thesis, Chalmers University of Technology, Gothenburg, Sweden.
Damaske, P. 1971. Head-related two-channel stereophony with loudspeaker reproduction, Journal of the Acoustical Society of America (4, pt. 2): 1109{1115.
Desper, S. 1995. Automatic stereophonic manipulation system and apparatus for
image enhancement, U.S. patent no. 5,412,731.
Duda, R. and Martens, W. 1997. Range-dependence of the HRTF for a spherical
head, Proceedings of the IEEE Workshop on Applications of Signal Processing
to Audio and Acoustics, New Paltz, New York.
Duda, R. and Martens, W. 1998. Range-dependence of the HRTF of a spherical
head, Journal of the Acoustical Society of America (5): 3048{3058.
102

139

97

30

37

50

104

172

BIBLIOGRAPHY

Duda, R. O., Avendano, C. and Algazi, V. R. 1999. An adaptable ellipsoidal head

model for the interaural time dierence, Proceedings of the IEEE International
Conference on Acoustics, Speech, and Signal Processing, Vol. 1, Phoenix, AZ,
USA, pp. 965{968.
Evans, M. J., Angus, J. A. S. and Tew, A. I. 1998. Analyzing head-related transfer
functions using surface spherical harmonics, Journal of the Acoustical Society
of America
(4): 2400{2411.
Fahy, F. J. 1995. Sound Intensity, 2nd edn, E & FN SPON.
Flanagan, J. L. 1960. Analog measurements of sound radiation from the mouth,
Journal of the Acoustical Society of America (12): 1613{1620.
Flanagan, J. L. 1972. Speech Analysis, Synthesis and Perception, 2nd edn,
Springer-Verlag, Berlin. pp. 38-41.
Fletcher, N. H. and Rossing, T. D. 1991. The Physics of Musical Instruments,
Springer-Verlag, New York, USA.
Foster, S., Wenzel, E. and Taylor, R. 1991. Real-time synthesis of complex acoustic environments, Proceedings of the IEEE Workshop on Applications of Signal
Processing to Audio and Acoustics, Mohonk Mountain House, New Paltz, New
York.
Friedlander, B. and Porat, B. 1984. The modied Yule-Walker method of ARMA
spectral estimation, IEEE Trans. Aerospace Electronic Syst.
(2): 158{
173.
Gaik, W. 1993. Combined evaluation of interaural time and intensity dierences:
Psychoacoustic results and computer modeling, Journal of the Acoustical Society of America (1): 98{110.
Gardner, W. G. 1995. Transaural 3-D audio, MIT Media Lab Perceptual Computing, technical report 342.
Gardner, W. G. 1997. 3-D Audio Using Loudspeakers, PhD thesis, MIT Media
Lab, p. 153.
Gardner, W. G. 1998a. 3-D Audio Using Loudspeakers, Kluwer Academic Publishers, Boston.
Gardner, W. G. 1998b. Reverberation algorithms, in: M. Kahrs and K. Brandenburg (eds), Applications of Digital Signal Processing to Audio and Acoustics,
Kluwer Academic Publishers, Norwell, MA, USA, chapter 3, pp. 85{131.
104

32

AES-20

94

173

BIBLIOGRAPHY

Gardner, W. G. and Martin, K. D. 1994. HRTF measurements of a KEMAR

dummy-head microphone, MIT Media Lab Perceptual Computing, technical
report 280.
Gardner, W. G. and Martin, K. D. 1995. HRTF measurements of a KEMAR,
Journal of the Acoustical Society of America (6): 3907{3908.
Gerzon, M. 1973. Periphony: with-height sound reproduction, Journal of the
Audio Engineering Society (1/2): 2{10.
Gerzon, M. 1992. Panpot laws for multispeaker stereo, the 92nd Convention of
the Audio Engineering Society, preprint 3309, Vienna, Austria.
Gilkey, R. and Anderson, T. 1997. Binaural and Spatial Hearing in Real and
Virtual Environments, Lawrence Erlbaum Associates, Mahwah, New Jersey,
p. 795.
Golub, G. H. and Loan, C. F. V. 1996. Matrix Computations, 3 edn, Princeton
University Press, The Johns Hopkins University Press, p. 694.
Gutknecht, M., Smith, J. O. and Trefethen, L. N. 1983. The CaratheodoryFejer (CF) method for recursive digital lter design, IEEE Transactions on
Acoustics, Speech, and Signal Processing (6): 1417{1426.
Hanninen, R. 1999. LibR { An Object-Oriented Software Architecture for Realtime
Sound and Kinematics, Helsinki University of Technology. Lic.Tech. Thesis.
Hartung, K. and Raab, A. 1996. Ecient modeling of head-related transfer
functions, Acta Acustica (suppl. 1): S88.
Hiipakka, D. G. J., Hanninen, R., Ilmonen, T., Napari, H., Lokki, T., Savioja,
L., Huopaniemi, J., Karjalainen, M., Tolonen, T., Valimaki, V., Valimaki, S.
and Takala, T. 1997. Virtual orchestra performance, Visual Proceedings of
SIGGRAPH'97, Los Angeles, p. 81.
Huopaniemi, J. 1997. Modeling of Human Spatial Hearing in the Context of Digital Audio and Virtual Environments, Helsinki University of Technology, Department of Electrical and Communications Engineering, Laboratory of Acoustics an d Audio Signal Processing, p. 101. Lic.Tech. Thesis.
Huopaniemi, J. and Karjalainen, M. 1996a. Comparison of digital lter design
methods for 3-D sound, Proc. IEEE Nordic Signal Processing Symposium, Espoo, Finland, pp. 131{134.
Huopaniemi, J. and Karjalainen, M. 1996b. HRTF lter design based on auditory criteria, Proc. Nordic Acoustical Meeting (NAM'96), Helsinki, Finland,
pp. 323{330.
97

21

31

82

174

BIBLIOGRAPHY

Huopaniemi, J. and Karjalainen, M. 1997. Review of digital lter design and

implementation methods for 3-D sound, Presented at the 102nd Convention of
the Audio Engineering Society, preprint 4461, Munich, Germany.
Huopaniemi, J. and Riederer, K. 1998. Measuring and modeling the eect of
distance in head-related transfer functions, Proceedings of the joint meeting of
the Acoustical Society of America and the International Congress on Acoustics,
Seattle, WA, USA, pp. 2083{2084.
Huopaniemi, J. and Smith, J. O. 1999. Spectral and time-domain preprocessing
and the choice of modeling error criteria for binaural digital lters, Proceedings of the AES 16th International Conference on Spatial Sound Reproduction,
Rovaniemi, Finland, pp. 301{312.
Huopaniemi, J., Karjalainen, M. and Valimaki, V. 1995. Physical models of
musical instruments in real-time binaural room simulation, Proceedings of the
International Congress on Acoustics, Trondheim, Norway, pp. 447{450.
Huopaniemi, J., Karjalainen, M., Valimaki, V. and Huotilainen, T. 1994. Virtual
instruments in virtual rooms { a real-time binaural room simulation environment for physical modeling of musical instruments, Proceedings of the International Computer Music Conference, Aarhus, Denmark, pp. 455{462.
Huopaniemi, J., Kettunen, K. and Rahkonen, J. 1999a. Measurement and modeling techniques for directional sound radiation from the mouth, Proceedings of
the IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, Mohonk Mountain House, New Paltz, New York.
Huopaniemi, J., Savioja, L. and Karjalainen, M. 1997. Modeling of re ections
and air absorption in acoustical spaces: a digital lter design approach, Proceedings of the IEEE Workshop on Applications of Signal Processing to Audio
and Acoustics, Mohonk Mountain House, New Paltz, New York.
Huopaniemi, J., Savioja, L. and Takala, T. 1996. DIVA virtual audio reality
system, Proceedings of the International Conference on Auditory Display, Palo
Alto, California, USA.
Huopaniemi, J., Zacharov, N. and Karjalainen, M. 1998. Objective and subjective evaluation of head-related transfer function lter design, Presented at the
105th Audio Engineering Society Convention, preprint 4805 (invited paper),
San Francisco, CA, USA.
Huopaniemi, J., Zacharov, N. and Karjalainen, M. 1999b. Objective and subjective evaluation of head-related transfer function lter design, Journal of the
Audio Engineering Society (4): 218{239.
47

BIBLIOGRAPHY

175

Imai, S. 1983. Cepstral analysis synthesis on the Mel frequency scale, Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal
Processing, Boston, MA, USA, pp. 93{96.
ISO 1985. ISO 354. Acoustics | Measurement of sound absorption in a reverberation room.
ISO 1989. ISO 8253-1. Acoustics | Audiometric test methods | Part 1: Basic
pure tone air and bone conduction threshold audiometry, pp. 587{599.
ISO 1993. ISO 9613-1. Acoustics | Attenuation of sound during propagation
outdoors | Part 1: Calculation of the absorption of sound by the atmosphere,
pp. 377{385.
ISO 1996. ISO 10534-1. Acoustics | Determination of sound absorption coefcient and impedance in impedance tubes | Part 1: Method using standing
wave ratio.
ISO/IEC 1997. ISO/IEC JTC/SC24 IS 14772-1. Information technology { Computer graphics and image processing - The Virtual Reality Modeling Language
(VRML97).
ISO/IEC 1999. ISO/IEC JTC1/SC29/WG11 IS 14496 (MPEG-4). Information
technology { Coding of audiovisual objects.
ITU-R 1994a. Recommendation BS.1116. Methods for the subjective assessment
of small impairments in audio systems including multichannel sound systems.
Geneva.
ITU-R 1994b. Recommendation BS.775-1. Multichannel stereophonic sound system with and without accompanying picture. International Telecommunications Union Radiocommunication Assembly, Geneva.
ITU-R 1997. Recommendation BS.1116-1. Methods for the subjective assessment of small impairments in audio systems including multichannel sound
systems. International Telecommunications Union Radiocommunication Assembly, Geneva.
Jeress, L. A. 1948. A place theory of sound localization, J. Comp. Physiol.
Psych. : 468{486.
Jenison, R. L. 1995. A spherical basis function neural network for pole-zero
modeling of head-related transfer functions, Proceedings of the IEEE Workshop
on Applications of Signal Processing to Audio and Acoustics, New Paltz, New
York.
61

176

BIBLIOGRAPHY

Jot, J.-M. 1999. Real-time spatial processing of sounds for music, multimedia
and interactive human-computer interfaces, Multimedia Systems : 55{69.
Jot, J.-M. and Chaigne, A. 1996. Method and system for articial spatialisation
of digital audio signals, U.S. patent no. 5,491,754.
Jot, J.-M., Jullien, J.-P. and Warusfel, O. 1998a. Method to simulate the acoustical quality of a room and associated audio-digital processor, U.S. patent no.
5,812,674.
Jot, J.-M., Larcher, V. and Warusfel, O. 1995. Digital signal processing issues
in the context of binaural and transaural stereophony, Presented at the 98th
Convention of the Audio Engineering Society, preprint 3980, Paris, France.
Jot, J.-M., Wardle, S. and Larcher, V. 1998b. Approaches to binaural synthesis,
Presented at the 105th Audio Engineering Society Convention, preprint 4861
(invited paper), San Francisco, CA, USA.
Juhl, P. M. 1993. The Boundary Element Method for Sound Field Calculations,
Ph.D. thesis, The Acoustics Laboratory, Technical University of Denmark, report 55.
Kale, I., Gryka, J., Cain, G. D. and Beliczynski, B. 1994. FIR lter order reduction: balanced model truncation and Hankel-norm optimal approximation,
IEE Proc.-Vis. Image Signal Process.
(3): 168{174.
Karjalainen, M. 1990. DSP software integration by object-oriented programming
- a case study of QuickSig, IEEE ASSP Magazine pp. 21{31.
Karjalainen, M., Harma, A. and Laine, U. K. 1997a. Realizable warped IIR
lters and their properties, Proceedings of the IEEE International Conference
on Acoustics, Speech, and Signal Processing, Munich, Germany.
Karjalainen, M., Harma, A., Laine, U. and Huopaniemi, J. 1997b. Warped lters
and their audio applications, Proceedings of the IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, Mohonk Mountain House,
New Paltz, New York.
Karjalainen, M., Huopaniemi, J. and Valimaki, V. 1995. Direction-dependent
physical modeling of musical instruments, Proceedings of the International
Congress on Acoustics, Vol. 3, Trondheim, Norway, pp. 451{454.
Karjalainen, M., Piirila, E., Jarvinen, A. and Huopaniemi, J. 1999. Comparison
of loudspeaker equalization methods based on DSP techniques, Journal of the
Audio Engineering Society (1/2): 14{31.
7

141

47

177

BIBLIOGRAPHY

Kendall, G. 1995. A 3-D sound primer: directional hearing and stereo reproduction, Computer Music Journal (4): 23{46.
Kendall, G. S. and Martens, W. L. 1984. Simulating the cues of spatial hearing in natural environments, Proceedings of the International Computer Music
Conference, Paris, France, pp. 111{125.
Kendall, G. S. and Rodgers, C. A. P. 1982. The simulation of three-dimensional
localization cues for headphone listening, Proceedings of the International Computer Music Conference.
Kirkeby, O. and Nelson, P. A. 1998. Digital lter design for virtual source imaging
systems, Presented at the 104th Convention of the Audio Engineering Society,
preprint 4688, Amsterdam, The Netherlands.
Kirkeby, O., Nelson, P. and Hamada, H. 1997. The stereo dipole - binaural
sound reproduction using two closely spaced loudspeakers, Presented at the
102nd Convention of the Audio Engineering Society, preprint 4463, Munich,
Germany.
Kirkeby, O., Nelson, P. and Hamada, H. 1998. The stereo dipole - binaural
sound reproduction using two closely spaced loudspeakers, Journal of the Audio
Engineering Society (5): 387{395.
Kirkeby, O., Rubak, P., Johansen, L. G. and Nelson, P. A. 1999. Implementation
of cross-talk cancellation networks using warped FIR lters, Proceedings of the
AES 16th International Conference on Spatial Sound Reproduction, Rovaniemi,
Finland, pp. 358{365.
Kistler, D. J. and Wightman, F. L. 1992. A model of head-related transfer
functions based on principal components analysis and minimum-phase reconstruction, Journal of the Acoustical Society of America (3): 1637{1647.
Klayman, A. 1988. Stereo enhancement system, U.S. patent no. 4,748,669.
Kleiner, M., Dalenback, B.-I. and Svensson, P. 1993. Auralization { an overview,
Journal of the Audio Engineering Society (11): 861{875.
Koenen, R. 1999. MPEG-4: Multimedia for our time, IEEE Spectrum pp. 26{33.
Koring, J. and Schmitz, A. 1993. Simplifying cancellation of cross-talk for playback of head-related recordings in a two-speaker system, Acustica : 221{
232.
Kotorynski, K. 1990. Digital binaural/stereo conversion and crosstalk canceling,
Proc. 89th Convention of the Audio Engineering Society, preprint 2949, Los
Angeles, USA.
19

46

91

41

179

178

BIBLIOGRAPHY

Krokstad, A., Strom, S. and Sorsdal, S. 1968. Calculating the acoustical room
response by the use of a ray tracing technique, Journal of Sound and Vibration
(1): 118{125.
Kuhn, G. F. 1977. Model for the interaural time dierences in the azimuthal
plane, Journal of the Acoustical Society of America (1): 157{167.
Kuhn, G. F. 1987. Physical acoustics and measurements pertaining to directional hearing, in: W. A. Yost and G. Gourevitch (eds), Directional Hearing,
Springer-Verlag, pp. 3{25.
Kulkarni, A. and Colburn, H. S. 1995a. Ecient nite-impulse-response lter
models of the head-related transfer function, Journal of the Acoustical Society
of America (5): 3278.
Kulkarni, A. and Colburn, H. S. 1995b. Innite-impulse-response lter models of
the head-related transfer function, Journal of the Acoustical Society of America
(5): 3278.
Kulkarni, A. and Colburn, H. S. 1997a. Finite-impulse-response models of the
head-related transfer-function. Submitted to J. Acoust. Soc. Am.
Kulkarni, A. and Colburn, H. S. 1997b. Innite-impulse-response models of the
head-related transfer-function. Submitted to J. Acoust. Soc. Am.
Kulkarni, A., Isabelle, S. K. and Colburn, H. S. 1995. On the minimum-phase approximation of head-related transfer functions, Proceedings of the IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, New Paltz,
New York.
Kulkarni, A., Isabelle, S. K. and Colburn, H. S. 1999. Sensitivity of human subjects to head-related transfer-function phase spectra, Journal of the Acoustical
Society of America
(5): 2821{2840.
Kulowski, A. 1985. Algorithmic representation of the ray tracing technique, Applied Acoustics (6): 449{469.
Kuttru, H. 1995. Sound eld prediction in rooms, Proc. 15th Int. Congr. Acoust.,
Vol. 2, Trondheim, Norway, pp. 545{552.
Kuttru, K. H. 1991. Room Acoustics, 3rd edn, Elsevier Science Publishers,
Essex, England.
Kuttru, K. H. 1993. Auralization of impulse responses modeled on the basis of
ray-tracing results, Journal of the Audio Engineering Society (11): 876{880.
8

62

97

97

105

18

41

179

BIBLIOGRAPHY

Kyriakakis, C. 1998. Fundamental and technological limitations of immersive

audio systems, Proceedings of the IEEE (5): 941{951.
Kyriakakis, C. and Holman, T. 1998. Video-based head tracking for improvements
in multichannel loudspeaker audio, Presented at the 105th Convention of the
Audio Engineering Society, preprint 4845, San Francisco, CA, USA.
Kyriakakis, C., Tsakalides, P. and Holman, T. 1999. Surrounded by Sound, IEEE
Signal Processing Magazine (1): 55{66.
Laakso, T. I., Valimaki, V., Karjalainen, M. and Laine, U. K. 1996. Splitting
the unit delay { tools for fractional delay lter design, IEEE Signal Processing
Magazine (1): 30{60.
Lahti, T. and Moller, H. 1996. The Sigyn Hall, Turku |- a concert hall of glass,
Proc. 1996 Nordic Acoustical Meeting, Helsinki, Finland, pp. 43{48.
Larcher, V. and Jot, J.-M. 1997. Techniques d'interpolation de ltres audionumeriques, Application a la reproduction spatiale des sons sur ecouteurs, Proc.
4th Congress on Acoustics, pp. 97{100.
Larcher, V., Jot, J.-M. and Vandernoot, G. 1998. Equalization methods in binaural technology, Presented at the 105th Convention of the Audio Engineering
Society, preprint 4858, San Francisco, CA, USA.
Lehnert, H. and Blauert, J. 1992. Principles of binaural room simulation, Applied
Acoustics (3-4): 259{291.
Lindemann, W. 1986. Extension of a binaural cross-correlation model by contralateral inhibition. I. Simulation of lateralization for stationary signals., Journal of the Acoustical Society of America (6): 1608{1622.
Lokki, T., Hiipakka, J., Hanninen, J., Ilmonen, T., Savioja, L. and Takala, T.
1999. Real-time audiovisual rendering and contemporary audiovisual art, Organized Sound.
Lowe, D., Willing, S., Gonnason, W. and Williams, S. 1995. Stereo enhancement
system, U.S. patent no. 5,440,638.
Lyon, R. and DeJong, R. 1995. Theory and Application of Statistical Energy
Analysis, 2nd edn, Butterworth-Heinemann, 313 Washington Street, Newton,
MA 02158.
MacCabe, C. J. and Furlong, D. J. 1991. Spectral stereo surround pan-pot,
Presented at the 90th Convention of the Audio Engineering Society, preprint
3067, Paris, France.
86

16

13

36

80

180

BIBLIOGRAPHY

Mackenzie, J., Huopaniemi, J., Valimaki, V. and Kale, I. 1997. Low-order modelling of head-related transfer functions using balanced model truncation, IEEE
Signal Processing Letters (2): 39{41.
Macpherson, E. A. 1991. A computer model of binaural localization for stereo
imaging measurement, Journal of the Audio Engineering Society (9): 604{
622.
Maher, R., Lindemann, E. and Barish, J. 1996. Old and new techniques for
articial stereophonic image enhancement, Presented at the 101st Convention
of the Audio Engineering Society, preprint 4371, Los Angeles, CA, USA.
Malham, D. and Myatt, A. 1995. 3-D sound spatialization using ambisonic techniques, Computer Music Journal (4): 58{70.
Marshall, A. H. and Meyer, J. 1985. The directivity and auditory impressions of
singers, Acustica : 130{140.
Martens, W. L. 1987. Principal components analysis and resynthesis of spectral
cues to perceived direction, Proceedings of the International Computer Music
Conference, pp. 274{281.
Martin, J., Maercke, D. V. and Vian, J.-P. 1993. Binaural simulation of concert
halls: a new approach for the binaural reverberation process, Journal of the
Acoustical Society of America (6): 3255{3264.
Mathworks 1994. MATLAB Signal Processing Toolbox. User's Manual.
McGrath, D. 1996. Method and apparatus for ltering an electronic environment with improved accuracy and eciency and short ow-through delay, U.S.
patent no. 5,502,747.
Mehrgardt, S. and Mellert, V. 1977. Transformation Characteristics of the External Human Ear, Journal of the Acoustical Society of America (6): 1567{1576.
Meyer, J. 1978. Acoustics and the Performance of Music, Verlag das Musikinstrument, Frankfurt/Main.
Middlebrooks, J. C. and Green, D. M. 1992. Observations on a principal components analysis of head-related transfer functions, Journal of the Acoustical
Society of America (1): 597{599. Letters to the Editor.
Mller, H. 1992. Fundamentals of binaural technology, Applied Acoustics (34): 171{218.
Mller, H., Srensen, M., Hammershi, D. and Jensen, C. 1995. Head-related
transfer functions of human subjects, Journal of the Audio Engineering Society
(5): 300{321.
4

39

19

58

94

61

92

36

43

181

BIBLIOGRAPHY

Mller, H., Srensen, M., Jensen, C. and Hammershi, D. 1996. Binaural technique: Do we need individual recordings?, Journal of the Audio Engineering
Society (6): 451{469.
Moore, B. C. 1981. Principal component analysis in linear systems: Controllability, observability, and model reduction, IEEE Trans. Automat. Contr.
Moore, B. C. J., Glasberg, B. R. and Baer, T. 1997. A model for the prediction
of thresholds, loudness, and partial loudness, Journal of the Audio Engineering
Society (4): 224{240.
Moore, B. C. J., Peters, R. W. and Glasberg, B. R. 1990. Auditory lter shapes
at low center frequencies, Journal of the Acoustical Society of America : 132{
140.
Moore, B. C. J., Peters, R. W. and Glasberg, B. R. 1996. A revision of Zwicker's
loudness model, Acta Acustica : 335{345.
Moorer, J. A. 1979. About this reverberation business, Computer Music Journal
(2): 13{28.
Morse, P. M. and Ingard, U. K. 1968. Theoretical Acoustics, Princeton University
Press, Princeton, New Jersey.
Naylor, G. and Rindel, J. 1994. Odeon Room Acoustics Program, Version 2.5,
User Manual, Technical Univ. of Denmark, Acoustics Laboratory, Publication
49, Denmark.
Nelson, P. A., Hamada, H. and Elliott, S. 1992. Adaptive inverse lters for
stereophonic sound reproduction, IEEE Transactions on Signal Processing
(7): 1621{1632.
Nelson, P. A., Ordu~na-Bustamante, F. and Hamada, H. 1995. Inverse lter design
and equalization zones in multichannel sound reproduction, IEEE Transactions
on Speech and Audio Processing (3): 185{192.
Nelson, P. A., Ordu~na-Bustamante, F. and Hamada, H. 1996a. Multichannel
signal processing techniques in the reproduction of sound, Journal of the Audio
Engineering Society (11): 973{989.
Nelson, P. A., Ordu~na-Bustamante, F., Engler, D. and Hamada, H. 1996b. Experiments on a system for the synthesis of virtual acoustic sources, Journal of
the Audio Engineering Society (11): 990{1007.
Olive, S. 1998. Subjective evaluation of 3-D sound based on two loudspeakers,
Presented at the 105th Audio Engineering Society Convention, San Francisco,
CA, USA.
44

45

88

82

40

44

44

182

BIBLIOGRAPHY

Oppenheim, A. V. and Schafer, R. W. 1989. Discrete-Time Signal Processing,

Prentice Hall, New Jersey.
Parks, T. and Burrus, C. 1987. Digital Filter Design, John Wiley&Sons, New
York.
Proakis, J. G. and Manolakis, D. G. 1992. Digital Signal Processing: Principles,
Algorithms, and Applications, 2nd . edn, Macmillan, p. 969.
Pulkki, V. 1997. Virtual sound source positioning using vector base amplitude
panning, Journal of the Audio Engineering Society (6): 456{466.
Pulkki, V., Karjalainen, M. and Huopaniemi, J. 1998. Analyzing virtual sound
sources using a binaural auditory model, Presented at the 104th Convention of
the Audio Engineering Society, preprint 4697, Amsterdam, The Netherlands.
Pulkki, V., Karjalainen, M. and Huopaniemi, J. 1999. Analyzing virtual sound
sources using a binaural auditory model, Journal of the Audio Engineering
Society (4): 204{217.
Rabinowitz, W. M., Maxwell, J., Shao, Y. and Wei, M. 1993. Sound localization
cues for a magnied head: implications from sound diraction about a rigid
sphere, Presence (2): 125{129.
Rasmussen, K. and Juhl, P. 1993. The eect of head shape on spectral stereo
theory, Journal of the Audio Engineering Society (3): 135{142.
Rayleigh, L. 1904. On the acoustic shadow of a sphere, Phil. Transact. Roy. Soc.
London
: 87{99.
Rayleigh, L. 1907. On our perception of sound direction, Philos. Mag. : 214{
232.
Riederer, K. 1998a. Repeatability analysis of HRTF measurements, Presented
at the 105th Convention of the Audio Engineering Society, preprint 4846, San
Francisco, CA, USA.
Riederer, K. A. J. 1998b. HRTF Measurements, Master's thesis, Helsinki University of Technology, Laboratory of Acoustics and Audio Signal Processing,
p. 134.
Roads, C. 1995. The Computer Music Tutorial, The MIT Press, Cambridge,
Massachusetts, USA.
Runkle, P. R., Blommer, M. A. and Wakeeld, G. H. 1995. A comparison of head
related transfer function interpolation methods, Proceedings of the IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, Institute of
the Electrical and Electronic Engineers, New Paltz, New York, pp. 88{91.
45

47

41

203A

13

183

BIBLIOGRAPHY

Ryan, C. and Furlong, D. 1995. Eects of headphone placement on headphone

equalization for binaural reproduction, Presented at the 98th Convention of the
Audio Engineering Society, preprint 4009, Paris, France.
Sandvad, J. and Hammershi, D. 1994a. Binaural auralization. Comparison of
FIR and IIR lter representation of HIRs, Presented at the 96th Convention of
the Audio Engineering Society, preprint 3862, Amsterdam, The Netherlands.
Sandvad, J. and Hammershi, D. 1994b. What is the most ecient way of representing HTF lters?, Proc. NORSIG'94, pp. 174{178.
Savioja, L. 1999. Modeling Techniques for Virtual Acoustics, PhD thesis, Helsinki
University of Technology, Laboratory of Telecommunications Software and
Multimedia, Espoo, Finland.
Savioja, L., Backman, J., Jarvinen, A. and Takala, T. 1995. Waveguide mesh
method for low-frequency simulation of room acoustics, Proceedings of the International Congress on Acoustics, Vol. 2, Trondheim, Norway, pp. 637{640.
Savioja, L., Huopaniemi, J., Lokki, T. and Vaananen, R. 1997. Virtual environment simulation - Advances in the DIVA project, Proceedings of the International Conference on Auditory Display, Palo Alto, California, USA.
Savioja, L., Huopaniemi, J., Lokki, T. and Vaananen, R. 1999. Creating interactive virtual acoustic environments, Journal of the Audio Engineering Society
(9): 675{705.
Scheirer, E. D., Vaananen, R. and Huopaniemi, J. 1999. AudioBIFS: Describing
audio scenes with the MPEG-4 multimedia standard, IEEE Trans. Multimedia
(3): 237{250.
Schroeder, M. and Atal, B. 1963. Computer simulation of sound transmission in
rooms, IEEE Conv. Record, pt. 7 pp. 150{155.
Schroeder, M. R. 1962. Natural-sounding articial reverberation, Journal of the
Audio Engineering Society (3): 219{223.
Schroeder, M. R. 1970. Digital simulation of sound transmission in reverberant
spaces, Journal of the Acoustical Society of America (2 (Part 1)): 424{431.
Schroeder, M. R. 1973. Computer models for concert hall acoustics, Am. J.
Physics : 461{471.
Semidor, C. and Couthon, L. 1998. Sound sources and simulation softwares
in room acoustics, Proceedings of the joint meeting of the Acoustical Society
of America and the International Congress on Acoustics, Seattle, WA, USA,
pp. 363{364.
47

10

47

41

184

BIBLIOGRAPHY

Shinn-Cunningham, B., Lehnert, H., Kramer, G., Wenzel, E. and Durlach, N.

1997. Auditory displays, in: R. Gilkey and T. Anderson (eds), Binaural and
spatial hearing in real and virtual environments, Lawrence Erlbaum Associates,
Dahwah, New Jersey, pp. 611{663.
Slaney, M. 1993. An ecient implementation of the Patterson-Holdsworth lter
bank, Apple Technical Report 35. Apple Computer, Inc.
Smith, J. O. 1983. Techniques for Digital Filter Design and System Identication with Application to the Violin, PhD thesis, Electrical Engineering Dept.,
Stanford University (CCRMA). Available as CCRMA Technical Report STANM-14.
Smith, J. O. 1991. Viewpoints on the history of digital synthesis, Proceedings of
the International Computer Music Conference, Montreal, Canada, pp. 1{10.
Smith, J. O. 1997. Acoustic modeling using digital waveguides, in: C. Roads,
S. T. Pope, A. Piccialli and G. De Poli (eds), Musical Signal Processing, Swets
& Zeitlinger, Lisse, the Netherlands, chapter 7, pp. 221{264.
Smith, J. O. 1998. Principles of digital waveguide models of musical instruments,
in: M. Kahrs and K. Brandenburg (eds), Applications of Digital Signal Processing to Audio and Acoustics, Kluwer Academic Publishers, Norwell, MA,
USA, chapter 10, pp. 417{466.
Smith, J. O. and Abel, J. S. 1995. The Bark bilinear transform, Proceedings of the IEEE Workshop on Applications of Signal Processing
to Audio and Acoustics, New Paltz, New York. Available online at
http://www-ccrma.stanford.edu/~jos/.
Smith, J. O. and Abel, J. S. 1999. The Bark and ERB bilinear transforms,
IEEE Transactions on Speech and Audio Processing. Accepted for publication.
Available online at http://www-ccrma.stanford.edu/~jos/.
Spandock, F. 1934. Ann. Physik V : 345.
Steiglitz, K. 1996. A Digital Signal Processing Primer, Addison-Wesley, Menlo
Park, CA.
Stone, H. and Sidel, J. L. 1993. Sensory Evaluation Practices, 2nd edn, Academic
Press, pp. 221{222.
Strube, H. W. 1980. Linear prediction on a warped frequency scale, Journal of
the Acoustical Society of America pp. 1071{1076.
Sugiyama, K. and Irii, H. 1991. Comparison of the sound pressure radiation from
a prolate spheroid and the human mouth, Acustica : 271{276.
20

73

185

BIBLIOGRAPHY

Svensson, U. P., Andersson, R. and Vanderkooy, J. 1997. A new interpretation of

the Biot-Tolstoy edge diraction model. Submitted to Journal of the Acoustical
Society of America.
Swaminathan, V., Vaananen, R., Fernando, G., Singer, D. and Belknap, W.
1999. MPEG-4 version 2 Committee Draft 14496-1 (Systems) ISO/IEC
JTC1/SC29/WG11, Seoul, Doc. W2739.
Takala, T., Hanninen, R., Valimaki, V., Savioja, L., Huopaniemi, J., Huotilainen,
T. and Karjalainen, M. 1996. An integrated system for virtual audio reality,
Presented at the 100th Convention of the Audio Engineering Society, preprint
4229, Copenhagen, Denmark.
Takeuchi, T., Nelson, P. and Kirkeby, O. 1997. Robustness of the performance
of the stereo dipole to misaligment of head position, Presented at the 102nd
Convention of the Audio Engineering Society, preprint 4464, Munich, Germany.
Theile, G. 1991. On the naturalness of two-channel stereo sound, Journal of the
Audio Engineering Society (10): 761{767.
Thurlow, W. R., Mangels, J. W. and Runge, P. S. 1967. Head movements during
sound localization, Journal of the Acoustical Society of America (2): 489{
493.
Tolonen, T., Valimaki, V. and Karjalainen, M. 1998. Evaluation of Modern Sound
Synthesis Methods, Helsinki University of Technology, Lab. of Acoustics and
Audio Signal Processing, Report 47, p. 114.
Toshiyuki, T., Tomohiro, M. and Yasuhisa, O. 1994. Surround signal processing
apparatus, European patent application no. 94305664.8.
Vaananen, R. and Huopaniemi, J. 1999. Spatial presentation of sound in scene
description languages, Presented at the 106th Convention of the Audio Engineering Society, preprint 4921, Munich, Germany.
Vaananen, R., Valimaki, V., Huopaniemi, J. and Karjalainen, M. 1997. Ecient
and parametric reverberator for room acoustics modeling, Proceedings of the
International Computer Music Conference, Thessaloniki, Greece, pp. 200{203.
Valimaki, V. 1995. Discrete-Time Modeling of Acoustic Tubes Using Fractional
Delay Filters, Dr. Tech. thesis, Helsinki University of Technology, Lab. of
Acoustics and Audio Signal Processing, Report 37.
Valimaki, V. and Laakso, T. I. 1998. Suppression of transients in variable recursive digital lters with a novel and ecient cancellation method, IEEE
Transactions on Signal Processing : 3408{3414.
39

42

46

186

BIBLIOGRAPHY

Valimaki, V., Huopaniemi, J., Karjalainen, M. and Janosy, Z. 1996. Physical

modeling of plucked string instruments with application to real-time sound
synthesis, Journal of the Audio Engineering Society (5): 331{353.
Valio, H. and Huopaniemi, J. 1996. Improved stereo enhancement, Finnish patent
no. FI962181.
Vercoe, B. L., Gardner, W. G. and Scheirer, E. D. 1998. Structured audio:
Creation, transmission, and rendering of parametric sound representations,
Proceedings of the IEEE (5): 922{940.
Vian, J.-P. and Martin, J. 1992. Binaural room acoustics simulation: practical
uses and applications, Applied Acoustics (3-4): 293{305.
Walsh, M. J. and Furlong, D. J. 1995. Improved spectral stereo head model,
Presented at the 99th Convention of the Audio Engineering Society, preprint
4128, New York, NY, USA.
Ward, D. and Elko, G. 1998. Optimum loudspeaker spacing for robust crosstalk
cancellation, Proceedings of the IEEE International Conference on Acoustics,
Speech, and Signal Processing, Vol. 1, Seattle, WA, USA, pp. 3541{3544.
Ward, D. and Elko, G. 1999. Eect of loudspeaker position on the robustness of
acoustic crosstalk cancellation, IEEE Signal Processing Letters (5): 106{108.
Watanabe, Y., Tokunou, H. and Hamada, H. 1996. Subjective investigation of
a new sound reproduction system (stereo dipole), Journal of the Acoustical
Society of America
(4, pt.2): 2809.
Wenzel, E. and Foster, S. 1993. Perceptual consequences of interpolating headrelated transfer functions during spatial synthesis, Proceedings of the IEEE
Workshop on Applications of Signal Processing to Audio and Acoustics, New
Paltz, New York.
Wenzel, E. M. 1992. Spatial sound and sonication. Presented in the International Conference on Auditory Display (ICAD'92). Also published in \Auditory
Display: Sonication, Audication, and Auditory Interface, SFI Studies in the
Sciences of Complexity", Proc. XVIII, edited by G. Kramer, Addison-Wesley,
1994).
Wenzel, E. M., Arruda, M., Kistler, D. J. and Wightman, F. L. 1993. Localization using nonindividualized head-related transfer functions., Journal of the
Acoustical Society of America (1): 111{123.
Widrow, B. and Stearns, S. 1985. Adaptive Signal Processing, Prentice-Hall,
p. 474.
44

86

36

100

94

187

BIBLIOGRAPHY

Wightman, F. L. and Kistler, D. J. 1989. Headphone simulation of free-eld

listening. I: stimulus synthesis, Journal of the Acoustical Society of America
(2): 858{867.
Wightman, F. L. and Kistler, D. J. 1992. The dominant role of low-frequency interaural time dierences in sound localization, Journal of the Acoustical Society
of America : 1648{1661.
Wightman, F. L. and Kistler, D. J. 1999. Resolution of front-back ambiguity
in spatial hearing by listener and source movement, Journal of the Acoustical
Society of America
(5): 2841{2853.
Woodworth, R. S. and Schlosberg, R. 1954. Experimental Psychology, Holt, Rinehard and Winston, pp. 349{361.
Wu, S. and Putnam, W. 1997. Minimum perceptual spectral distance FIR lter
design, Proceedings of the IEEE International Conference on Acoustics, Speech,
and Signal Processing, Vol. 1, Munich, Germany, pp. 447{450.
Yost, W. A. and Gourevitch, G. 1987. Directional Hearing, Springer-Verlag,
p. 305.
Zacharov, N. and Huopaniemi, J. 1999. Results of a round robin subjective
evaluation of virtual home theatre systems, Presented at the 107th Convention
of the Audio Engineering Society, preprint 5067, New York, USA.
Zacharov, N., Huopaniemi, J. and Hamalainen, M. 1999. Round robin subjective evaluation of virtual home theatre systems at the AES 16th international
conference, Proceedings of the AES 16th International Conference on Spatial
Sound Reproduction, Rovaniemi, Finland, pp. 544{556.
Zwicker, E. and Fastl, H. 1990. Psychoacoustics: Facts and Models, SpringerVerlag, Heidelberg, Germany, p. 354.
85

91

105

Index
error norms, 95
Chebyshev norm, 95
Hankel norm, 96
least-squares norm, 95
fractional delay, 71
frequency warping, 88
geometrical room acoustics, 57
head-related transfer function, 48, 67
distance dependency, 79
interpolation, 104
modeling, 85
preprocessing, 92
head-tracking, 158
image-source method, 37
interaural level dierence, 68, 77
interaural time dierence, 68, 71
elevation dependence, 75
material absorption, 58
minimum-phase reconstruction, 69
MPEG, 21, 27
MPEG-4, 40
multichannel, 30, 31, 158
objective evaluation, 106
Parseval's relation, 95
physical modeling, 44
principal components analysis, 85
principle of reciprocity, 48
room acoustics, 32
geometrical, 34
room impulse response rendering

3-D sound, 30
acoustical material absorption, 58
air absorption, 62
ambisonics, 32
auditory display, 20
auditory smoothing, 87
auditory weighting, 88
auralization, 20
balanced model truncation, 98
Bark, 86
binaural, 20, 30
modeling and reproduction, 67
room impulse response, 34
binaural auditory model, 107
error criteria, 108
binaural lter design, 90
balanced model truncation, 98
FIR methods, 96
IIR methods, 97
crosstalk canceling, 30, 139
decorrelation of virtual sources, 148
digital lter design
frequency sampling, 96
directivity, 45
directional ltering, 46
set of elementary sources, 47
sound radiation from the mouth,
48
DIVA, 22
duplex theory, 20, 77
equivalent rectangular bandwidth, 86
ERB, 86
188

INDEX

direct, 35
parametric, 35
perceptual modeling, 35
physical modeling, 35
singular value decomposition, 99
sound source
directivity, 45
modeling, 43
natural audio, 43
physical modeling, 44
synthetic audio, 44
source-medium-receiver concept, 19,
28
spherical head model, 49
stereo dipole, 140, 158
stereo widening, 145, 151
subjective evaluation, 106
SVD, 99
talking heads, 44
transaural, 20, 139
vector base amplitude panning, 31
virtual acoustic display, 21
virtual acoustics, 20, 27
implementation, 32
modeling concepts, 27
virtual home theater, 21, 161
virtual loudspeaker, 145
lter design, 154
virtual surround, 21, 161
VRML, 21, 27, 40
warped lters, 100
unwarping, 89
warped FIR, 100
warped IIR, 102
warped shuer lter, 156, 163
warping coecient, 89

189